Deep learning has been a blessing and a curse for classification tasks. The performance gains this paradigm has brought are enormous, but they come at the cost of vast and often unrealistic amounts of labelled data. Conventionally, training a traditional classification model requires at least some data samples for all target classes, and deep learning models significantly amplify this issue. Collecting training instances of every class is not always easy, especially in fine-grained image classification (Li et al., 2018, 2019c; Xie et al., 2019b), and therefore much attention has been given to zero-shot learning (ZSL) algorithms as a solution (Akata et al., 2015a; Changpinyo et al., 2016; Li et al., 2019b; Yang et al., 2016; Liu et al., 2019; Qin et al., 2017; Gao et al., 2020; Shen et al., 2020; Xie et al., 2019a). ZSL expands the classifiers beyond the seen classes with abundant data to unseen classes without enough image samples.
Generally, ZSL can be categorised into two groups based on how the semantic features are leveraged. These are embedding-based methods and generative methods. Within the embedding-based methods, there are three mainstream methods to learn embeddings: (1) visual-to-semantic mappings; (2) semantic-to-visual mappings; or (3) latent embedding spaces. Class labels predicted by scoring the compatibility of the visual and semantic features (Akata et al., 2013, 2015b; Changpinyo et al., 2016; xie，Li Liu，Fan Zhu，Fang Zhao，Zheng Zhang，Yazhou Yao，Jie Qin，Ling Shao, 2019). By contrast, generative methods deal with ZSL by augmenting data with semantic information and using generative models to ”imagine” visual features of the unseen classes. The ZSL task can then be converted into a supervised classification problem (Xian et al., 2018a; Zhu et al., 2018).
Generative ZSL is currently considered to be the state-of-the-art. These approaches allow visual features to be generated for unseen classes with intra-class diversity while retaining inter-class discrimination. A range of techniques have been used to generalize the knowledge learned from seen classes to unseen classes. These include generative adversarial nets (GANs) (Goodfellow et al., 2014)
, variational autoencoders (VAEs)(Kingma and Welling, 2013) and some hybrid variants centred around complicated network design (Huang et al., 2019; Li et al., 2019a; Huang et al., 2019; Chen et al., 2020; Xian et al., 2019). All these methods derive the semantic information needed to generate features from external sources, such as Wikipedia.
The methods based on generative models either synthesise global visual features extracted from full images, as shown in Figure 1(a) or they synthesise features from multiple patches per image. By contrast, patch generation methods first identify various semantic patches in each image, e.g., head, tail, breast, then they extract the relevant feature from each patch, as shown in Figure 1(b). It is important to note, however, that both patch generation and global generation methods generate all the features for all patches via a single generator.
Despite their promising performance, these approaches have two major limitations in scenarios where fine-grained class distinctions are required. The first is the inability to deal with background noise. The second is the lack of a way to weight the importance of some features over others.
Background Noise (Yu et al., 2018): With global generation methods, the extracted visual representations inevitably include some of the background information, which is noise and distracts attention from the discriminative aspects of the local patches. This is particularly problematic in fine-grained classification because objects in two similar classes may only be distinguishable by a few subtle semantic difference.
Weighting for Importance: Most patch generation methods do not include a mechanism for weighting the discriminative power of each patch. Rather, they let the neural network make its own obscure decisions about which features are important and which are not. However, directly leveraging our own knowledge that some features are highly discriminative, while others are shared and therefore entirely irrelevant can bring great benefits to ZSL most notably, improved accuracy and reduced time complexity. For instance, Crackling Geese and Canada Geese are almost identical, except for the length of the neck. Therefore, in a task to distinguish Crackling geese from Canadian geese, learning to recognise features other than the neck, like the breast or tail, would not only be a waste of time, it would also increase the computational overhead. Focusing on the length of the neck, however, would guide the model toward greater accuracy while reducing the cost of training at the same time.
To address these issues, we designed a novel divide-and-conquer framework, called multi-patch generative adversarial nets (MPGAN), that separately synthesises local features for each semantic patch and classifies unseen classes with a novel voting strategy. The framework is illustrated in Figure 1(c). To alleviate the negative impact of background noise, each image is decomposed into multiple patches based on semantic descriptions, which largely eliminates the irrelevant background information. Within each patch, a paired generator and discriminator synthesise the visual feature associated with the patch according to the semantic information. Because each patch has its own GAN, as opposed to the single GAN used by other patch generation methods, the features synthesised by our framework are of inherently better quality in comparison.
To capture the most relevant patterns, a classifier is constructed for each patch that can recognise fine-grained classes. The final classification is made through an ensemble vote on the probability outputs of all the classifiers. Weight assignments based on the discriminative power of the semantic patches ensure the most important patches are brought to prominence. For example, in a task to classify images of birds, the patch containing the ”head” feature may be more discriminative than the ”tail” patch; therefore, the head would be weighted more highly than the tail. This weighting, i.e., the level of discrimination, is determined by an attention mechanism that computes both the inter-class and intra-class distances of the features to the class centroid for each patch and dividing the two. With the weightings allocated and the ensemble vote complete, the class labels with the highest confidence in the stacked probability output are taken as the final prediction.
Overall, our contributions can be briefly summarised as follows:
We propose a divide-and-conquer framework, called MPGAN, that divides an image into multiple semantic patches. Local visual features for each patch are then generated by their own dedicated GAN to improve accuracy with fine-grained ZSL classification tasks. To the best of our knowledge, this is the first attempt to isolate patches and generate visual features with dedicated GANs in generative ZSL.
To exploit prominent local patterns, each patch has its own classifier. Further, a novel attention mechanism calculates weights for each patch based on its discriminative power. These weights help the classification model focus on the patches with more distinctive patterns while skipping those patches with common or identical visual characteristics.
Comprehensive experiments and in-depth analyses with two real-world benchmark datasets demonstrate state-of-the-art accuracy by the MPGAN framework in both zero-shot image recognition and zero-shot retrieval.
The rest of the paper is organised as follows. We briefly review related work in Section 2. MPGAN is presented in Section 3, followed by the experiments in Section 4. Lastly, Section 5 concludes the paper.
2. Related Work
2.1. Generative ZSL
A number of generative methods have been proposed for generating visual features in ZSL. GAZSL (Zhu et al., 2018) is a carefully designed generative model based on the improved WGAN (Gulrajani et al., 2017) that synthesises realistic visual features from noisy Wikipedia articles. These new visual centroids provide complementary supervision for regulating the visual feature distribution with inter-class discrimination. CIZSL (Elhoseiny and Elfeki, 2019) directly generates unseen classes by making adjustments, i.e., deviations, to seen classes with the help of a parametrised entropy measure. CADA-VAE (Schonfeld et al., 2019) is a dual VAE model that finds an intermediate hidden feature space between the visual and attribute spaces. CANZSL (Chen et al., 2020) leverages a cycle architecture by translating synthesised visual features into semantic information. A cycle-consistent loss is then applied between the ground truth and the synthesised semantic information. GDAN (Huang et al., 2019) incorporates a flexible metric in the model’s discriminator to measure the similarity of features from different modalities. LisGAN (Li et al., 2019a) considers the multi-view nature of different images and regularises each generated sample to be close to at least one fundamentally representative sample.
Of these methods, GAZSL (Zhu et al., 2018), CIZSL (Elhoseiny and Elfeki, 2019), CANZSL (Chen et al., 2020) all extract visual features from multiple patches like our framework. However, the patch features in these methods are concatenated as global visual representations, then handled in the same way as global generation methods, thus, neglecting the distinct discriminative information in different patches. By contrast, MPGAN computes a weight for each patch, which results in substantially greater accuracy.
2.2. Fine-grained Image Generation
A recent trend, especially with multi-stage approaches, is to progressively synthesise vivid fine-grained images at the pixel level using multiple generative modules (Zhang et al., 2017, 2018; Xu et al., 2018; Chen and Luo, 2019). The process with StackGAN (Zhang et al., 2017), for example, involves sketching the basic shape and colours of an object given a text description in Stage 1, then using the text descriptions and the results from the previous stage as inputs to generate high-resolution images with photo-realistic details in Stage 2. alignDRAW (Mansimov et al., 2015) generates images from a natural language description and considers fine-grained information by iteratively drawing image patches on a canvas, while attending to the relevant words in the caption. TextureGAN (Xian et al., 2018b) generates plausible fine-grained textures in deep image synthesis with a local patch sampler.
Unlike these pixel-level image generation methods, ZSL for classification tasks only involves feature-level generation. The aim is to generate an image that can be visually interpreted while generating visual features that can directly provide discrimination power for better classification outcomes. In pursuit of better discrimination among generated visual features, MPGAN therefore synthesises multiple discriminative patch features in parallel.
2.3. Attention Mechanisms in ZSL
In recent years, attention mechanisms have boosted the performance of many deep learning tasks. In ZSL, attention helps to break through the limitations imposed by noisy global features. More specifically, it allows the classification model to attend to the most discriminative areas, effectively eliminating background noise.
In this vein, S2GA (Yu et al., 2018) is an end-to-end framework with a stacked semantically-guided attention mechanism that progressively guides the visual features toward generated attention weights for different local features. Notably, S2GA integrates both the global visual features and the weighted local discriminant features to represent an image by allocating different weights for different local regions based on their relevance to the semantic features of the class. SGMA (Zhu et al., 2019) is a localised semantically-guided multi-attention mechanism directed toward global features. A multi-attention loss encourages compact and diverse attention distribution by applying geometric constraints over attention maps. The AREN model (Xie et al., 2019a) discovers multiple semantic parts of images guided by an attention mechanism and the compatibility loss. The model is also coupled with a paralleled network branch to guarantee more stable semantic transfer from the perspective of second-order collaboration.
The conventional attention techniques incorporated into these approaches typically have a high time complexity during training and fail to directly attend to the most critical areas. Hence, MPGAN incorporates a novel discrimination-based attention mechanism that operates as an offline algorithm with pre-computed attention weights.
This section begins with the problem formulation and notations. MPGAN is outlined next, followed by the discrimination-based attention mechanism and the ensembled voting strategy.
3.1. Problem Formulation and Notations
Consider a dataset where the classes are split into two sets - a seen set and an unseen set .Thus, the seen class labels in and the unseen class labels in are disjoint, i.e., = . The semantic descriptions of the seen and unseen classes are available as text embeddings in the semantic space, denoted as and , respectively. Each class label in either the seen or unseen set corresponds to a text embedding or . Given labeled training instances in , where denotes the visual features of the n-th image and is the corresponding class label. Each consists of patches, of which each is represented as , where . The aim of this task is to build a model that can accurately classify visual features for unseen classes by learning to generate visual features from semantic information. There are testing instances in , where are the visual features of the -th unseen testing sample and is the unseen class label.
3.2. Generating Visual Patches
The MPGAN framework is shown in Figure 2. The basic specialist module for generating each visual patch is a generative adversarial net (GAN) consisting of a discriminator network D and a generator network G. Each net is independent and generates visual features corresponding to its allocated patch from semantic information. For example, one of the GANs, consisting of and , may be allocated to the head patches and so would be trained to only synthesise the head features. Unlike existing patch generation methods where one GAN generates all the features for all the patches (e.g. head, back, wing, etc.) (Zhu et al., 2018; Chen et al., 2020; Elhoseiny and Elfeki, 2019), using multiple independent GANs dedicated to generating just the features for one patch results in far more robust visual features.
We selected WGAN-GP (Gulrajani et al., 2017) as the basic generative block because this method introduces a Wasserstein distance and a gradient penalty specifically to prevent the problem of model collapse during training a common problem with GANs. Further, given that the generated visual features need to be conditioned on the corresponding semantic information, we incorporated Zhu et al.’s GAZSL approach (Zhu et al., 2018) with its baseline settings into the framework to generate a GAN for each semantic patch. GAZSL is an extended version of an adversarial net with an additional classification loss and a regularisation term to increase inter-class discrimination. Each generator
takes the Term Frequency-Inverse Document Frequency (TF-IDF) representations extracted from the semantic descriptions as its input. However, not all the text in a description is necessarily relevant to a specific patch. For instance, descriptions of the ”Tail” and ”Back” features are not relevant to the ”Head” feature. This information is treated as noise and removed before synthesising the target feature with a denoising module. The denoising module could be either a fully connected layer or Principal Component Analysis (PCA). For convenience, we have provided a comparison between the two in the experiments section.
Once the denoised text representations have been prepared, they are concatenated with some random noise
sampled from a Gaussian distribution. These concatenated representations are then used as the input for two subsequent fully connected layers, followed by the activation functions. The objective of the generatorof the -th patch is defined as
where represents the Wasserstein loss for the -th patch generation, and is a classification loss to ensure the inter-class discrimination of the synthesised visual features. the parameter for the -th patch generator, and is the parameter for the discriminator. The last term is the visual pivot regularisation that aims to push the generated samples towards the corresponding visual centre of cluster and patch in the Euclidean space, denoted as . is the weight for the visual pivot in the -th patch.
The discriminator for the
-th patch then takes the synthesised visual features as input to calculate both the probability that the features are real as well as the classification loss. Specifically, each discriminator consists of three fully connected layers. The first layer encodes the visual features of the patch, followed by a ReLU activation function. The other two fully connected layers take the encoded visual features and calculate the discriminator loss and the classification loss, respectively. The discriminator loss is a basic loss that distinguishes how realistic the input patch features are. The classification loss is to distinguish the synthesised visual features between classes, which is needed to be able to explicitly classify each feature into a corresponding class. The objective function of the discriminatorformulated as
where the first two represents the classification losses - one for generated patch features and the other for the real patch features. is the gradient penalty used to stabilize GAN training, and is a Lipschitz constraint, , where
is a linear interpolation between the fake featuresand the real features in the -th patch, and is the weight of the gradient penalty in the -th patch. The last term is the Wasserstein loss for the synthesised and real input visual features.
3.3. Ensemble Prediction for Unseen Patches
With the model trained to generate visual features from semantic information, the next step is to generate realistic visual features for the unseen classes from the semantic information. Once complete, supervised classifiers can be trained to predict labels for each patch.
However, because the overarching recognition task covers all patches, there needs to be a strategy to improve global classification performance by integrating the output probabilities from each patch into the final prediction. A typical approach is a voting strategy over all the individual classification results, but we want to give priority to the more discriminative patches so, as an initial implementation, we incorporated a traditional self-attention mechanism as a weighting method. However, preliminary experiments showed that self-attention did not significantly affect the results of our framework. Considering that the patches are already determined during generation, we then replaced the self-attention mechanism with a straightforward weighted voting technique, reducing the complexity of the framework in the process.
3.3.1. Discriminative Attention
With the evidence that self-attention had merit, we next need to tackle the issue of weighting the importance of each patch. Inspired by the weighted voting strategy and the failed attention mechanism above, we develop an alternative attention mechanism to weight the importance of each patch based on its discrimination power during the process of predicting labels for unseen classes. These discrimination-based attention weights are computed in a pre-processing step by leveraging the intra- and inter-class difference between the seen classes in the training set for each patch. The computation process is shown in Algorithm 1. By way of summary, first the cluster centroids for each seen class in the -th patch are identified. The intra-class distance for is then determined by the sum of the Euclidean distance between each visual sample and its associated cluster centroid, and the inter-class distance is measured according to the distance of the nearest sample from each other class to the class centroid . Next, the inter-class/intra-class distance ratio is computed from an element-wise division of and . Lastly, the mean value of the discrimination ratio for all the classes is yielded as the attention weight for each patch.
With the calculation above, the intra-class distance summarises the dispersity of the class distributions, whereas the inter-class distance reflects how distant the class is from other class distributions. As the inter intra-class distance ratio is the quotient of the inter-class distance and the intra-class distance, the less disperse the classes, the higher the ratio. And, given that the ratio of all classes is averaged for each patch, the higher the average ratio, the more discriminative the patch.
3.3.2. Ensemble Prediction
One softmax classifier is trained for each patch . The patch features inputs into the classifiers first pass through a fully connected layer with the number of unseen classes as the output dimension:
where are the synthesised unseen -th patch features, and and are the parameters of the fully connected layer. Each classifier is independently trained to derive the probability distribution over the unseen classes, where , and is the total number of unseen classes . Making a prediction for a patch is done by simply choosing the class index with the highest probability:
However, a single patch cannot accurately recognize an object, we then integrate the probability distributions of the various patches to make the ensemble prediction together with the attention weights that we computed above. Since an object cannot be accurately recognised from a single patch, the probability distributions of various patches are integrated and an ensemble prediction is made together with the attention weights computed above.
The probability distribution of each patch is multiplied with the corresponding attention weight ; thus, the scaled probability distribution can be calculated with
is the element-wise product. Each probability value in the vectoris multiplied with the scalar to increase or reduce the probability according to the attention weight. Element-wise addition is then performed on the scaled probabilities for each patch :
The final prediction is made by picking the highest probability from the weighted probability distribution :
. The descriptive statistics are provided in Table 1. CUB is a small-scale dataset with 11,788 bird images spanning 200 fine-grained species, whereas NAB is a significantly larger dataset with 48,562 images and 1,011 categories. The NAB dataset also has a class hierarchy with 456 parent nodes and 555 leaf nodes. Elhoseiny et al.(Elhoseiny et al., 2017) annotated the two datasets with semantic descriptions from Wikipedia articles. However, descriptions of some species were missing so, in the NAB dataset, these species were merged with other classes, leaving 404 annotated classes. There are two suggested splitting strategies for testing classifications tasks with these datasets. The first is the super-category-shared splitting strategy (SCS), where unseen categories with relatively high relevance to seen categories are chosen to share the same super-class. The other strategy is the super-category-exclusive strategy (SCE), where all classes belong to the same super-class but are split into either seen or unseen categories. Intuitively, zero-shot recognition performance should be better with the SCS strategy than with SCE.
|Dataset||TF-IDF Dim||SCS||SCE||#Images||#Patches||Patch Dim|
Since the ultimate goal of ZSL is highly accurate classification, most algorithms operate directly on ready-to-use features extracted from the images, which, as a bonus, results in significantly reduced computation time. The visual features of the CUB and NAB datasets are extracted from VPDEnet following the settings in (Elhoseiny et al., 2017). The CUB dataset returned seven visual patches: ”head”, ”back”, ”belly”, ”breast”, ”leg”, ”wing” and ”tail”. NAB lacks the ”leg” patch, so there are only six semantic patches. All patches have 512 dimensions. Notably, the superiority of VPDE-net as a global feature extractor has not yet been proven, so we also choose to test the widely-used CNN framework ResNet101 as a comparison. The results of this experiment are given in Section 4.4. All the raw text annotations are tokenised into words with punctuation removed, then processed into TF-IDF representations. CUB’s representations have 7,551 dimensions, while NAB’s have 13,217 (due to differences in the word counts for each dataset). This level of dimensionality is very high so, to reduce the computation cost, we filter out the irrelevant features and embed the representations into a lower-dimensional space in our method.
4.2. Evaluation Metrics and Comparison Methods
The average per-class Top-1 accuracy was used as the evaluation criteria, formulated as:
where is the number of testing classes. A correct prediction is defined as the highest probability of all candidate classes.
We choose six state-of-the-art algorithms for comparison: MCZSL (Akata et al., 2016), ZSLNS (Qiao et al., 2016), ZSLPP (Elhoseiny et al., 2017), GAZSL (Zhu et al., 2018) , S2GA (Yu et al., 2018), CANZSL (Chen et al., 2020), CIZSL (Elhoseiny and Elfeki, 2019). MCZSL directly uses partial annotations as strong supervision to extract CNN representations of the semantic information in the test phase. GAZSL and ZSLPP simply generate features based on the detected semantic information during both training and testing. S2GA uses semantically-guided attention to identify relevant features, which are then used to progressively generate an attention map for weighting the importance of different local regions. CANZSL is an extension of GAZSL based on cycle architecture that reveals semantic descriptions from visual features, yielding a significant performance improvement. CIZSL introduces a learning signal inspired by the literature on creativity. The unseen space is explored with imagined class-descriptions and careful deviations from the visual features generated from seen classes are encouraged while allowing knowledge transfer from seen to unseen classes. It is worth mentioning that the basic GAN module in our framework is simpler than the one in CANZSL. As mentioned, CANZSL is based on a cycle architecture designed to reveal semantic information with high accuracy. Therefore, for a fair comparison with GAZSL, the root of CANZSL, we do not add extra components to the GANs, such as a reverse generator like CANZSL.
4.3. Zero-Shot Learning Results
Our first experiment test general ZSL classification for each of the methods using the two recommended splitting strategies. We run each experiment multiple times, reporting the average classification accuracy for each dataset and each strategy. The results, shown in Table 2, attest that MPGAN was the most accurate of all the methods with both datasets and both strategies. On CUB, the performance improvement over the next best method, CANZSL, is 2.4% (SCS) and 1.1% (SCE) over CIZSL. On NAB, the improvement is 1.8% (SCS) and 0.3% (SCE) over S2GA. We made several other observations from the results as follows. (1) Accuracy with the SCE splitting strategy is significantly lower than with SCS, which indicates that the relevance of the unseen classes to the seen classes probably has a significant influence over the model’s generalisation ability among semantic and visual modalities. (2) Even though S2GA uses both global visual features and those from patches to generate the attention feature map, MPGAN was more accurate using the patch features alone. (3) MPGAN delivers a performance improvement on the CUB dataset of 4.5% (SCS) and 5.2% (SCE) over GAZSL and 5.7% (SCS) and 1.4% (SCE) on NAB. This is significant because GAZSL can be thought of as a basic version of our MPGAN framework. Overall, the results show a substantial improvement over the current state-of-the-arts, verifying MPGAN’s accuracy with ZSL classification tasks.
4.4. Comparison with Global Generation
The next set of experiments is designed to evaluate the quality of MPGAN’s generated patch and global features. We first fine-tune a ResNet101 model on a training set consisting of the seen classes from CUB. We then use the trained model without the last fully connected layer to extract 2048-dimensional visual features of both the seen and unseen classes. With TF-IDF representations, the global visual features only reach around 10.5% accuracy with SCS and 4.3% with SCE. Other ZSL methods that consider global visual features are capable of much better performance – for example, (Li et al., 2019a; Xian et al., 2019; Schonfeld et al., 2019). This unsatisfactory result is mainly because the semantic attributes in the dataset are manually annotated by domain experts and are therefore very clean. However, taking that amount of time to label the instances would seldom be practical in zero-shot settings. Hence, we simply extract the semantic information from Wikipedia articles.
4.5. Ablation Study
In this ablation study, we test different variants of MPGAN’s overall architecture. These included: MP-only generating patch features with global visual centroids; MP-MC-only generating features from multiple patches with multiple centroids; MP-MC-MCIs-only generating features from multiple patches with multiple centroids and predicting the labels via a single classifier; and MPGAN the complete method. The MP-only variant generates visual features for each patch separately, but a global visual centroid is applied to an ensemble of the generated visual patches. MP-only outperforms the GAZSL baseline, which generates visual features for all patches via a single GAN. This result indicates that parallel multi-patch generation is effective. The MP-MC-only variant tests how the centroids affect performance. The results of this experiment confirm that the centroids contribute to the discriminative power of the synthesised visual features. We also find that the voting strategy with multiple classifiers for each patch prediction, as opposed to MP-MC-MCIs-only with a single classifier, proved to be helpful. Further, there is an obvious improvement when attention weights are incorporated into the voting strategy. The specific results for each method appear in Table 3, showing that, as more components are added with each variant, performance gradually improves. Hence, each component makes a contribution to the accuracy of the final prediction.
4.6. Noise Suppression
Given the TF-IDF representations of text descriptions are high-dimensional (7,551 for CUB and 13,217 for NAB), the time complexity of training/inference can be relatively high if denoising operations involve the entire network. Hence, this experiment is designed to test whether pre-processing the text representations into a lower-dimensional feature space using PCA would deliver the same performance as dynamic denoising through a fully connected layer in the overall model. As there are only 200 classes in the CUB dataset, 200 principal components are the maximum that can be preserved. Obviously, reducing the dimensionality from 7,551 to 200 will result in the loss of some non-trivial information. In contrast, the 404 classes of the NAB dataset mean that significantly more information can be preserved. The performance comparisons between the two noise suppression methods are given in Table 2.
4.7. Parameter Sensitivity
In a series of experiments, we test the impact of main four hyper-parameters on accuracy. Varying the number of generated samples from 100 to 600 in steps of 50, we find little impact on accuracy, as shown in Figure 3(a). Overall, accuracy reaches its peak at around 200 samples for each unseen class on the NAB dataset and 350 on CUB. The number of denoised dimensions also has a relatively small impact. Figure 3(b) shows the results of denoising the TF-IDF representations at various levels of dimensionality through a fully connected layer before feeding them into the generative model. The highest accuracy for both datasets occurs at around 1000-dimensions, which would likely apply to any bird recognition task. However, the peak dimensions for other tasks would need to be confirmed in future research. The effect of different numbers of training epochs is shown in Figure 3(c). Overall, MPGAN proves to be very stable. After an initial climb, accuracy plateaued at around 500 iterations and only begins to overfit the seen classes at 2000 iterations. The last and perhaps most important sensitivity test is the influence of the different patches. The results in Figure 3(d), indicate that MPGAN is the least accurate when the head patch is excluded on both datasets. In a t-SNE analysis, Figure 4 confirms the discrimination of this particular patch. However, despite the importance of the head patch, the best performance occurs when every patch is considered, which demonstrates that each patch is critical.
4.8. t-SNE Analysis
The t-SNE visualisation in Figure 4 shows the extracted visual features and the synthesised visual features for the different semantic classes with the CUB dataset. The features included in the plot are randomly picked from unseen classes. It is clear that each patch has a different level of discriminative power. For example, the head, as discussed in the previous experiment, is extremely discriminative, as is the leg according to the ground truth. However, relying on the head patch alone will only result in an accuracy of about 35%. Since the framework is designed to push the generated samples towards the visual centroid of each class during training, the synthesised features are intra-class discriminative. However, due to the lack of discrimination in some patches, the distributions of the synthesised features for some classes significantly overlap. This issue illustrates the reasoning behind the discrimination-based attention mechanism, which mitigates the problem by assigning different weights based on the discriminative power of the patch.
4.9. Zero-shot Retrieval
To further verify the effectiveness of the MPGAN framework, we also perform the zero-shot retrieval task. In zero-shot retrieval, images of unseen classes must be classified and returned based only on semantic information. We choose mean average precision (mAP) as the evaluation metric, where Precision is calculated as the percentage of correct images retrieved for the class. Table 4 shows the accuracy of each method when retrieving 25%, 50% and 100% of the images for all unseen classes, and Figure 5 illustrates an example of the results for four randomly-chosen classes. The images shown are the five nearest images to the class. The results in Table 4 show MPGAN with the highest accuracy at the 25% and 50% levels on the CUB dataset, and at the 25% level with NAB. The first column of Figure 5 shows a typical image of the bird species. The rest of the columns show the nearest five images in order. Images with green boxes are correctly classified; red boxes are incorrect.
In this paper, we presents a generative paradigm for ZSL that breaks images into a series of fine-grained patches to reveal subtle visual distinctions in features. Called multi-patch generative adversarial networks (MPGAN), the framework decomposes object classification into one classification sub-task per patch. An attention-based voting mechanism then considers the subtle difference between fine-grained patches among classes. An extensive suite of experiments show that MPGAN is a more accurate solution for zero-shot image recognition than seven state-of-the-art approaches with competitive accuracy in zero-shot retrieval. In future work, we intend to generalise MPGAN to other nuanced subject matter, such as flowers or dogs.
- Akata et al. (2016) Zeynep Akata, Mateusz Malinowski, Mario Fritz, and Bernt Schiele. 2016. Multi-cue zero-shot learning with strong supervision. In CVPR. 59–68.
et al. (2013)
Zeynep Akata, Florent
Perronnin, Zaid Harchaoui, and Cordelia
Label-embedding for attribute-based
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 819–826.
- Akata et al. (2015a) Zeynep Akata, Florent Perronnin, Zaid Harchaoui, and Cordelia Schmid. 2015a. Label-embedding for image classification. TPAMI 38, 7 (2015), 1425–1438.
- Akata et al. (2015b) Zeynep Akata, Scott Reed, Daniel Walter, Honglak Lee, and Bernt Schiele. 2015b. Evaluation of output embeddings for fine-grained image classification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2927–2936.
- Changpinyo et al. (2016) Soravit Changpinyo, Wei-Lun Chao, Boqing Gong, and Fei Sha. 2016. Synthesized classifiers for zero-shot learning. In CVPR. 5327–5336.
- Chen et al. (2020) Zhi Chen, Jingjing Li, Yadan Luo, Zi Huang, and Yang Yang. 2020. CANZSL: Cycle-Consistent Adversarial Networks for Zero-Shot Learning from Natural Language. In WACV.
- Chen and Luo (2019) Zhi Chen and Yadan Luo. 2019. Cycle-Consistent Diverse Image Synthesis from Natural Language. In 2019 IEEE International Conference on Multimedia & Expo Workshops (ICMEW). IEEE, 459–464.
- Elhoseiny and Elfeki (2019) Mohamed Elhoseiny and Mohamed Elfeki. 2019. Creativity Inspired Zero-Shot Learning. In Proceedings of the IEEE International Conference on Computer Vision. 5784–5793.
- Elhoseiny et al. (2017) Mohamed Elhoseiny, Yizhe Zhu, Han Zhang, and Ahmed Elgammal. 2017. Link the head to the” beak”: Zero shot learning from noisy text description at part precision. In CVPR. 6288–6297.
- Gao et al. (2020) Rui Gao, Xingsong Hou, Jie Qin, Jiaxin Chen, Li Liu, Fan Zhu, Zhao Zhang, and Ling Shao. 2020. Zero-VAE-GAN: Generating Unseen Features for Generalized and Transductive Zero-Shot Learning. IEEE Transactions on Image Processing 29 (2020), 3665–3680.
- Goodfellow et al. (2014) Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial nets. In NeurlPS. 2672–2680.
- Gulrajani et al. (2017) Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. 2017. Improved training of wasserstein gans. In Advances in neural information processing systems. 5767–5777.
- Huang et al. (2019) He Huang, Changhu Wang, Philip S Yu, and Chang-Dong Wang. 2019. Generative Dual Adversarial Network for Generalized Zero-shot Learning. In CVPR. 801–810.
- Kingma and Welling (2013) Diederik P Kingma and Max Welling. 2013. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013).
- Li et al. (2019a) Jingjing Li, Mengmeng Jin, Ke Lu, Zhengming Ding, Lei Zhu, and Zi Huang. 2019a. Leveraging the Invariant Side of Generative Zero-Shot Learning. In CVPR.
- Li et al. (2019b) Jingjing Li, Mengmeng Jing, Ke Lu, Lei Zhu, Yang Yang, and Zi Huang. 2019b. Alleviating Feature Confusion for Generative Zero-shot Learning. In Proceedings of the 27th ACM International Conference on Multimedia. 1587–1595.
- Li et al. (2018) Jingjing Li, Lei Zhu, Zi Huang, Ke Lu, and Jidong Zhao. 2018. I read, i saw, i tell: Texts assisted fine-grained visual classification. In Proceedings of the 26th ACM international conference on Multimedia. 663–671.
- Li et al. (2019c) Zhengming Li, Zheng Zhang, Jie Qin, Zhao Zhang, and Ling Shao. 2019c. Discriminative fisher embedding dictionary learning algorithm for object recognition. IEEE transactions on neural networks and learning systems 31, 3 (2019), 786–800.
- Liu et al. (2019) Yang Liu, Jishun Guo, Deng Cai, and Xiaofei He. 2019. Attribute Attention for Semantic Disambiguation in Zero-Shot Learning. In Proceedings of the IEEE International Conference on Computer Vision. 6698–6707.
- Mansimov et al. (2015) Elman Mansimov, Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. 2015. Generating images from captions with attention. arXiv preprint arXiv:1511.02793 (2015).
- Qiao et al. (2016) Ruizhi Qiao, Lingqiao Liu, Chunhua Shen, and Anton Van Den Hengel. 2016. Less is more: zero-shot learning from online textual documents with noise suppression. In CVPR. 2249–2257.
- Qin et al. (2017) Jie Qin, Li Liu, Ling Shao, Fumin Shen, Bingbing Ni, Jiaxin Chen, and Yunhong Wang. 2017. Zero-shot action recognition with error-correcting output codes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2833–2842.
- Schonfeld et al. (2019) Edgar Schonfeld, Sayna Ebrahimi, Samarth Sinha, Trevor Darrell, and Zeynep Akata. 2019. Generalized zero-and few-shot learning via aligned variational autoencoders. In CVPR. 8247–8255.
- Shen et al. (2020) Yuming Shen, Jie Qin, and Lei Huang. 2020. Invertible Zero-Shot Recognition Flows. arXiv preprint arXiv:2007.04873 (2020).
- Van Horn et al. (2015) Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. 2015. Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. 595–604.
- Wah et al. (2011) Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. 2011. The caltech-ucsd birds-200-2011 dataset. (2011).
- Xian et al. (2018b) Wenqi Xian, Patsorn Sangkloy, Varun Agrawal, Amit Raj, Jingwan Lu, Chen Fang, Fisher Yu, and James Hays. 2018b. Texturegan: Controlling deep image synthesis with texture patches. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 8456–8465.
- Xian et al. (2018a) Yongqin Xian, Tobias Lorenz, Bernt Schiele, and Zeynep Akata. 2018a. Feature generating networks for zero-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition. 5542–5551.
- Xian et al. (2019) Yongqin Xian, Saurabh Sharma, Bernt Schiele, and Zeynep Akata. 2019. f-VAEGAN-D2: A feature generating framework for any-shot learning. In CVPR. 10275–10284.
- Xie et al. (2019a) Guo-Sen Xie, Li Liu, Xiaobo Jin, Fan Zhu, Zheng Zhang, Jie Qin, Yazhou Yao, and Ling Shao. 2019a. Attentive region embedding network for zero-shot learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 9384–9393.
- Xie et al. (2019b) Guo-Sen Xie, Zheng Zhang, Li Liu, Fan Zhu, Xu-Yao Zhang, Ling Shao, and Xuelong Li. 2019b. Srsc: Selective, robust, and supervised constrained feature representation for image classification. IEEE Transactions on Neural Networks and Learning Systems (2019).
- xie，Li Liu，Fan Zhu，Fang Zhao，Zheng Zhang，Yazhou Yao，Jie Qin，Ling Shao (2019) Guosen xie，Li Liu，Fan Zhu，Fang Zhao，Zheng Zhang，Yazhou Yao，Jie Qin，Ling Shao. 2019. Region Graph Embedding Network for Zero-Shot Learnin. In Proceedings of the IEEE International Conference on Computer Vision. 5784–5793.
- Xu et al. (2018) Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. 2018. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In CVPR. 1316–1324.
- Yang et al. (2016) Yang Yang, Yadan Luo, Weilun Chen, Fumin Shen, Jie Shao, and Heng Tao Shen. 2016. Zero-shot hashing via transferring supervised knowledge. In Proceedings of the 24th ACM international conference on Multimedia. 1286–1295.
Yu et al. (2018)
Yunlong Yu, Zhong Ji,
Yanwei Fu, Jichang Guo,
Yanwei Pang, Zhongfei Mark Zhang,
et al. 2018.
Stacked semantics-guided attention model for fine-grained zero-shot learning. InNeurIPS. 5995–6004.
- Zhang et al. (2017) Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. 2017. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In CVPR. 5907–5915.
- Zhang et al. (2018) Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. 2018. Stackgan++: Realistic image synthesis with stacked generative adversarial networks. TPAMI 41, 8 (2018), 1947–1962.
- Zhu et al. (2018) Yizhe Zhu, Mohamed Elhoseiny, Bingchen Liu, Xi Peng, and Ahmed Elgammal. 2018. A generative adversarial approach for zero-shot learning from noisy texts. In CVPR. 1004–1013.
- Zhu et al. (2019) Yizhe Zhu, Jianwen Xie, Zhiqiang Tang, Xi Peng, and Ahmed Elgammal. 2019. Semantic-Guided Multi-Attention Localization for Zero-Shot Learning. In NeurlPS. 14917–14927.