Large-scale Image Classification Research Articles

Remote Assistance for Bone-Fractured Patients using Deep Learning Models

ABSTRACT Remote diagnosis enables healthcare professionals to evaluate and diagnose patients from a distance using telecommunication technologies, enhancing healthcare delivery by improving accessibility, especially for those in remote or underserved areas. One of the significant sustainability challenges in remote medical diagnostics is offering timely assistance to vulnerable groups like the elderly, disabled, mentally impaired individuals, and wounded military personnel in combat zones. This becomes particularly difficult in emergencies when rapid analysis of medical records is needed, especially if the data is stored on secure blockchain networks. The proposed work addresses these challenges by deploying a comprehensive framework for large-scale analysis, utilizing both document and image classification for dual validation. It integrates advanced techniques such as Inception V3, VGG-16, VGG-19, RESNET-50, and Densenet-201 for bone fracture detection, with Inception V3 achieving the highest accuracy of 95.1%. In addition, a Document Classification Analysis (DCA) method is proposed, which automatically classifies the severity of fractures. Object detection techniques are also introduced for detecting minor fractures using region-based image segmentation, ensuring precise diagnosis even for subtle injuries. This pioneering integration of technologies provides a holistic solution for remote medical diagnostics.

Robust Structure-Aware Graph-based Semi-Supervised Learning: Batch and Recursive Processing

Graph-based semi-supervised learning plays an important role in large scale image classification tasks. However, the problem becomes very challenging in the presence of noisy labels and outliers. Moreover, traditional robust semi-supervised learning solutions suffers from prohibitive computational burdens thus cannot be computed for streaming data. Motivated by that, we present a novel unified framework robust structure-aware semi-supervised learning called Unified RSSL (URSSL) for batch processing and recursive processing robust to both outliers and noisy labels. Particularly, URSSL applies joint semi-supervised dimensionality reduction with robust estimators and network sparse regularization simultaneously on the graph Laplacian matrix iteratively to preserve the intrinsic graph structure and ensure robustness to the compound noise. First, in order to relieve the influence from outliers, a novel semi-supervised robust dimensionality reduction is applied relying on robust estimators to suppress outliers. Meanwhile, to tackle noisy labels, the denoised graph similarity information is encoded into the network regularization. Moreover, by identifying strong relevance of dimensionality reduction and network regularization in the context of robust semi-supervised learning (RSSL), a two-step alternative optimization is derived to compute optimal solutions with guaranteed convergence. We further derive our framework to adapt to large scale semi-supervised learning particularly suitable for large scale image classification and demonstrate the model robustness under different adversarial attacks. For recursive processing, we rely on reparameterization to transform the formulation to unlock the challenging problem of robust streaming-based semi-supervised learning. Last but not least, we extend our solution into distributed solutions to resolve the challenging issue of distributed robust semi-supervised learning when images are captured by multiple cameras at different locations. Extensive experimental results demonstrate the promising performance of this framework when applied to multiple benchmark datasets with respect to state-of-the-art approaches for important applications in the areas of image classification and spam data analysis.

PATH-13. LEARNED RESIZING WITH EFFICIENT TRAINING (LRET) FACILITATES IMPROVED PERFORMANCE OF LARGE-SCALE BRAIN TUMOR HISTOLOGY IMAGE CLASSIFICATION MODELS

Abstract BACKGROUND Histologic examination is vital in oncology research and diagnostics. The adoption of digital scanning of whole slide images (WSI) has created an opportunity to leverage deep learning-based image classification methods to enhance diagnosis and risk stratification. However, technical limitations prevent training and deployment of accurate comprehensive multiclass deep convolutional neural networks (DCNN) models for histopathology image classification. The input dimensions of DCNN architectures are small compared to the typical pathologist field of view, degrading performance by excluding important architectural features. Furthermore, data requirements for comprehensive models are sufficiently large to overwhelm the system memory during training. METHODS A method termed Learned Resizing with Efficient Training (LRET) was developed to address the main limitations of traditional histopathology classification model training. The LRET method couples efficient training techniques with image resizing to facilitate seamless integration of larger histology image patches into state-of-the-art classification models while preserving important structural information. The LRET method was coupled with two distinct resizing techniques to train three diverse histology image datasets using five different DCNN architectures. Performance metrics were compared on cross validation and hold out test sets. RESULTS LRET-trained models were flexible to multiple input patch dimensions and DCNN models. We demonstrated performance improvement across all datasets while significantly reducing the training time and resources over traditional methods. Using a large-scale, multiclass brain tumor classification dataset consisting of 74 distinct histopathologic classes, LRET-trained models outperformed existing methods by 15-28% in accuracy, yielding 94% accuracy for the best model. CONCLUSION The LRET method for DCNN training significantly enhances the performance of large-scale multiclass histopathology image classification. The implications of this work extend to broader applications within medical imaging and beyond, where efficient integration of high-resolution images into deep learning pipelines is paramount for driving advancements research and clinical practice.

Unsupervised Domain Adaptation with Contrastive Learning-Based Discriminative Feature Augmentation for RS Image Classification

High- and very high-resolution (HR, VHR) remote sensing (RS) images can provide comprehensive and intricate spatial information for land cover classification, which is particularly crucial when analyzing complex built-up environments. However, the application of HR and VHR images to large-scale and detailed land cover mapping is always constrained by the intricacy of land cover classification models, the exorbitant cost of collecting training samples, and geographical changes or acquisition conditions. To overcome this limitation, we propose an unsupervised domain adaptation (UDA) with contrastive learning-based discriminative feature augmentation (CLDFA) for RS image classification. In detail, our method first utilizes contrastive learning (CL) through a memory bank in order to memorize sample features and improve model performance, where the approach employs an end-to-end Siamese network and incorporates dynamic pseudo-label assignment and class-balancing strategies for adaptive domain joint learning. By transferring classification models trained on a source domain (SD) to an unlabeled target domain (TD), our proposed UDA method enables large-scale land cover mapping. We conducted experiments using a massive five billion-pixels dataset as the SD and tested the HR and VHR RS images of five typical Chinese cities as the TD and applied the method on the completely unlabeled world view 3 (WV3) image of Urumqi city. The experimental results demonstrate that our method excels in large-scale HR and VHR RS image classification tasks, highlighting the advantages of semantic segmentation based on end-to-end deep convolutional neural networks (DCNNs).

Detecting Post Editing of Multimedia Images using Transfer Learning and Fine Tuning

In the domain of general image forgery detection, a myriad of different classification solutions have been developed to distinguish a “tampered” image from a “pristine” image. In this work, we aim to develop a new method to tackle the problem of binary image forgery detection. Our approach builds upon the extensive training that state-of-the-art image classification models have undergone on regular images from the ImageNet dataset, and transfers that knowledge to the image forgery detection space. By leveraging transfer learning and fine tuning, we can fit state-of-the-art image classification models to the forgery detection task. We train the models on a diverse and evenly distributed image forgery dataset. With five models—EfficientNetB0, VGG16, Xception, ResNet50V2, and NASNet-Large—we transferred and adapted pre-trained knowledge from ImageNet to the forgery detection task. Each model was fitted, fine-tuned, and evaluated according to a set of performance metrics. Our evaluation demonstrated the efficacy of large-scale image classification models—paired with transfer learning and fine tuning—at detecting image forgeries. When pitted against a previously unseen dataset, the best-performing model of EfficientNetB0 could achieve an accuracy rate of nearly 89.7%.

Learning Deep Asymmetric Tolerant Part Representation

Categorization objects at a sub-ordinate level inevitably poses a significant challenge, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">i.e.</i> , inter-class difference is very subtle and only exists in a few key parts. Therefore, how to localize these key parts for discriminative visual categorization without requiring expensive pixel-level annotations becomes a core question. To that end, this paper introduces a novel asymmetry tolerant part segmentation network (ATP-Net). ATP-Net simultaneously learns to segment parts and identify objects in an end-to-end manner using only image-level category labels. Given the intrinsic asymmetry property of part alignment, a desirable learning of part segmentation should be capable of incorporating such property. Despite the efforts towards regularizing weakly supervised part segmentation, none of them consider this vital and intrinsic property, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">i.e.</i> , the spatial asymmetry of part alignment. Our work, for the first time, proposes to explicitly characterize the spatial asymmetry of part alignment for visual tasks. We propose a novel asymmetry loss function to guide the part segmentation by encoding the spatial asymmetry of part alignment, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">i.e.</i> , restricting the upper bound of how asymmetric those self-similar parts are to each other in the network learning. Via a comprehensive ablation study, we verify the effectiveness of the proposed ATP-Net in driving the network learning towards semantically meaningful part segmentation and discriminative visual categorization. Consistently superior/competitive performance are reported on 12 datasets covering crop cultivar classification, plant disease classification, bird/butterfly species classification, large-scale natural image classification, attribute recognition and landmark localization.

SAR-HUB: Pre-Training, Fine-Tuning, and Explaining

Since the current remote sensing pre-trained models trained on optical images are not as effective when applied to SAR image tasks, it is crucial to create sensor-specific SAR models with generalized feature representations and to demonstrate with evidence the limitations of optical pre-trained models in downstream SAR tasks. The following aspects are the focus of this study: pre-training, fine-tuning, and explaining. First, we collect the current large-scale open-source SAR scene image classification datasets to pre-train a series of deep neural networks, including convolutional neural networks (CNNs) and vision transformers (ViT). A novel dynamic range adaptive enhancement method and a mini-batch class-balanced loss are proposed to tackle the challenges in SAR scene image classification. Second, the pre-trained models are transferred to various SAR downstream tasks compared with optical ones. Lastly, we propose a novel knowledge point interpretation method to reveal the benefits of the SAR pre-trained model with comprehensive and quantifiable explanations. This study is reproducible using open-source code and datasets, demonstrates generalization through extensive experiments on a variety of tasks, and is interpretable through qualitative and quantitative analyses. The codes and models are open source.

EARNet: Error-Aware Reconstruction Network for no-reference image quality assessment

Deep learning-based no-reference image quality assessment (NRIQA) methods have demonstrated advanced performance. In this paper, a deep learning-based NRIQA method with strong error-aware and content-aware capabilities is proposed, which consists of Error-Aware Reconstruction Network (EARNet) module, Content Feature Extraction Network (CFENet) module, and Subjective Quality Regression Network (SQRNet) module. We first build a database to pre-train EARNet to obtain the ability to extract error features. For content features, CFENet pre-trained on large-scale image classification tasks is adopted to extract. The pre-trained EARNet and CFENet are serially connected with SQRNet so that the features received by SQRNet are both error-aware and content-aware. Extensive experimental results show that the proposed method achieves state-of-the-art performance on many well-known IQA databases. The robustness of the proposed method is verified on the large-scale Waterloo Exploration Database (WED), and its superiority is demonstrated by the group maximum differentiated (gMAD) competition game. Furthermore, we also verify that the proposed EARNet is highly extensible, which can further improve the performance of the existing deep learning-based NRIQA method.

Deep learning based data augmentation for large-scale mineral image recognition and classification

Vision-based mineral image recognition and classification is a proven solution for autonomous unmanned ore sorting. Although accurate identification can be achieved by training models offline using large-scale datasets, the lack of sufficient labeled images still limits the accessibility and exploration of high-performance deep learning models. To address the above issues, referring to the generative adversarial networks, three different deep learning-based mineral image data augmentation models are proposed in this work. The experimental results show that the proposed models can generate mineral images with high fidelity and have high similarity to the ground truth in terms of texture, color and shape. Compared with classic data augmentation methods, proposed ones can better optimize downstream sorting tasks: the accuracy of ResNet101, ResNet50, InceptionV3 and VGG19 is improved by 18.52%, 9.94%, 4.39% and 2.39%, respectively. Finally, this work also presents a monolithic three-stage system workflow for large-scale mineral image recognition and classification.

Content-based product image retrieval using squared-hinge loss trained convolutional neural networks

&lt;p&gt;&lt;span lang="EN-US"&gt;Convolutional neural networks (CNN) have proven to be highly effective in large-scale object detection and image classification, as well as in serving as feature extractors for content-based image retrieval. While CNN models are typically trained with category label supervision and softmax loss for product image retrieval, we propose a different approach for feature extraction using the squared-hinge loss, an alternative multiclass classification loss function. First, transfer learning is performed on a pre-trained model, followed by fine-tuning the model. Then, image features are extracted based on the fine-tuned model and indexed using the nearest-neighbor indexing technique. Experiments are conducted on VGG19, InceptionV3, MobileNetV2, and ResNet18 CNN models. The model training results indicate that training the models with squared-hinge loss reduces the loss values in each epoch and reaches stability in less epoch than softmax loss. Retrieval results show that using features from squared-hinge trained models improves the retrieval accuracy by up to 3.7% compared to features from softmax-trained models. Moreover, the squared-hinge trained MobileNetV2 features outperformed others, while the ResNet18 feature gives the advantage of having the lowest dimensionality with competitive accuracy.&lt;/span&gt;&lt;/p&gt;

Diverse Sample Generation: Pushing the Limit of Generative Data-Free Quantization.

Generative data-free quantization emerges as a practical compression approach that quantizes deep neural networks to low bit-width without accessing the real data. This approach generates data utilizing batch normalization (BN) statistics of the full-precision networks to quantize the networks. However, it always faces the serious challenges of accuracy degradation in practice. We first give a theoretical analysis that the diversity of synthetic samples is crucial for the data-free quantization, while in existing approaches, the synthetic data completely constrained by BN statistics experimentally exhibit severe homogenization at distribution and sample levels. This paper presents a generic Diverse Sample Generation (DSG) scheme for the generative data-free quantization, to mitigate detrimental homogenization. We first slack the statistics alignment for features in the BN layer to relax the distribution constraint. Then, we strengthen the loss impact of the specific BN layers for different samples and inhibit the correlation among samples in the generation process, to diversify samples from the statistical and spatial perspectives, respectively. Comprehensive experiments show that for large-scale image classification tasks, our DSG can consistently quantization performance on different neural architectures, especially under ultra-low bit-width. And data diversification caused by our DSG brings a general gain to various quantization-aware training and post-training quantization approaches, demonstrating its generality and effectiveness.

Class-Incremental Learning: Survey and Performance Evaluation on Image Classification.

For future learning systems, incremental learning is desirable because it allows for: efficient resource usage by eliminating the need to retrain from scratch at the arrival of new data; reduced memory usage by preventing or limiting the amount of data required to be stored - also important when privacy limitations are imposed; and learning that more closely resembles human learning. The main challenge for incremental learning is catastrophic forgetting, which refers to the precipitous drop in performance on previously learned tasks after learning a new one. Incremental learning of deep neural networks has seen explosive growth in recent years. Initial work focused on task-incremental learning, where a task-ID is provided at inference time. Recently, we have seen a shift towards class-incremental learning where the learner must discriminate at inference time between all classes seen in previous tasks without recourse to a task-ID. In this paper, we provide a complete survey of existing class-incremental learning methods for image classification, and in particular, we perform an extensive experimental evaluation on thirteen class-incremental methods. We consider several new experimental scenarios, including a comparison of class-incremental methods on multiple large-scale image classification datasets, an investigation into small and large domain shifts, and a comparison of various network architectures.

Extension of Divisible-Load Theory from Scheduling Fine-Grained to Coarse-Grained Divisible Workloads on Networked Computing Systems

The big data explosion has sparked a strong demand for high-performance data processing. Meanwhile, the rapid development of networked computing systems, coupled with the growth of Divisible-Load Theory (DLT) as an innovative technology with competent scheduling strategies, provides a practical way of conducting parallel processing with big data. Existing studies in the area of DLT usually consider the scheduling problem with regard to fine-grained divisible workloads. However, numerous big data loads nowadays can only be abstracted as coarse-grained workloads, such as large-scale image classification, context-dependent emotional analysis and so on. In view of this, this paper extends DLT from fine-grained to coarse-grained divisible loads by establishing a new multi-installment scheduling model. With this model, a subtle heuristic algorithm was proposed to find a feasible load partitioning scheme that minimizes the makespan of the entire workload. Simulation results show that the proposed algorithm is superior to the up-to-date multi-installment scheduling strategy in terms of achieving a shorter makespan of workloads when dealing with coarse-grained divisible loads.

Inducing semantic hierarchy structure in empirical risk minimization with optimal transport measures

The cross-entropy (CE) loss is arguably the most important empirical risk minimization objective for deep discriminative models for classification, and has achieved notable success in numerous applications. Though the CE loss is widely adopted, it essentially ignores the correlation between categories. For example, predicting a shepherd dog to husky is more acceptable than a tiger for the subsequent decision processes, while these two misclassifications result in the same CE loss. Therefore, the usually used CE loss does not incorporate the risk of misclassification of different categories, which can be measured by the distance between the predicted category and ground-truth category in a semantic hierarchical tree (SHT). In this work, to explicitly take the SHT-defined risk-aware inter-categorical correlation into consideration, by proposing a discrete optimal transport (DOT) training framework via configuring its ground distance matrix. We are able to predefine ground distance matrix in optimal transport measurement following a priori of hierarchical semantic risk. Specifically, the tree-induced error (TIE) on SHT is adopted as our ground distance matrix. Furthermore, it can be extended to its increasing function from the optimization perspective. In addition, we can also adaptively learn the matrix following an alternative optimization scheme. The semantic similarity in each level of a tree is integrated with the information gain. We demonstrated the effectiveness of our framework in several benchmarks of large-scale image classification with the semantic tree structure, and showed superior performance in a plug-and-play manner.11The code is available in: https://anonymous.4open.science/r/OTM-Neurocomputing/

Preface to the First Issue

Welcome to the first issue of Artificial Intelligence and Applications (AIA) Journal. Recently, we are seeing the rapid development of Artificial Intelligence (AI) field, evidenced by the fast-growing published papers and increased practical applications. Artificial intelligence is a discipline of science and technology for making intelligent machines to simulate human abilities in learning, perception, thinking, decision, behaviour and interaction. The technology is widely used in various areas in the society, including sensing data analysis and understanding, big data analytics, security surveillance, management and planning, manufacturing, education, medical care, robotics, unmanned driving, and so on. The artificial intelligence field was found in 1956, when the term “artificial intelligence” was coined and came into the open. Since then, many theories, methods and algorithms have been developed under the main paradigms of symbolism, connectionism and behaviourism. The large boom came in the latest decade. The milestones of successful applications of deep neural networks (DNNs) in large scale image classification in 2012 and computer Go game in 2016 drew the eye of AI researchers to the deep learning approach and made AI widely known to the society. The research is extended to all aspects of intelligence science and technology, and the application of AI is extended into all scenarios of our life. The published papers are quickly increasing in existing journals and conferences, and new journals and conferences are being established. In preparing this journal, we selected the name Artificial Intelligence and Applications (AIA) because both research and application are important to the field and both are fast developing. Real-life applications of AI are enabled in recent years because the developed methodologies have led to improved performance satisfying users’ needs. On the other hand, the extensive applications also pose scientific and technical challenges which attract researchers’ attention. Such interaction between research and application will continually drive the evolution of the field. The Artificial Intelligence and Applications (AIA) is a peer-reviewed journal publishing original contributions to the theory, methods and applications of artificial intelligence. Qualified papers are the most important factor to establish the fame and impact of the journal. We solicit papers reporting novel ideas, innovative technology or applications in the AI field. We thank the authors of all submitted papers for their great efforts in supporting our journal. Let’s grow up together with the AIA Journal!

Large-scale image classification with multi-perspective deep transfer learning

Most research efforts on image classification so far have been focused on medium-scale datasets. In addition, there exist other problems, such as difficulty in feature extraction and small sample size. In order to address above difficulties, this paper proposes a multi-perspective convolutional neural network model, which contains channel attention module and spatial attention module. The proposed modules derive attention graphs from channel dimension and spatial dimension respectively, then the input features are selectively learned according to the importance of the features. We explain how the gain in storage can be traded against a loss in accuracy and/or an increase in CPU cost. In addition, we give the interpretability of the model at multiple scales. Quantitative and qualitative experimental results demonstrate that the accuracy of our proposed model can be improved by up to 3.8% and outperforms the state-of-the-art methods.

One-shot Federated Learning on Medical Data using Knowledge Distillation with Image Synthesis and Client Model Adaptation.

One-shot federated learning (FL) has emerged as a promising solution in scenarios where multiple communication rounds are not practical. Notably, as feature distributions in medical data are less discriminative than those of natural images, robust global model training with FL is non-trivial and can lead to overfitting. To address this issue, we propose a novel one-shot FL framework leveraging Image Synthesis and Client model Adaptation (FedISCA) with knowledge distillation (KD). To prevent overfitting, we generate diverse synthetic images ranging from random noise to realistic images. This approach (i) alleviates data privacy concerns and (ii) facilitates robust global model training using KD with decentralized client models. To mitigate domain disparity in the early stages of synthesis, we design noise-adapted client models where batch normalization statistics on random noise (synthetic images) are updated to enhance KD. Lastly, the global model is trained with both the original and noise-adapted client models via KD and synthetic images. This process is repeated till global model convergence. Extensive evaluation of this design on five small- and three large-scale medical image classification datasets reveals superior accuracy over prior methods. Code is available at https://github.com/myeongkyunkang/FedISCA.

NASiam: Efficient Representation Learning using Neural Architecture Search for Siamese Networks

Siamese networks are one of the most trending methods to achieve self-supervised visual representation learning (SSL). Since hand labeling is costly, SSL can play a crucial part by allowing deep learning to train on large unlabeled datasets. Meanwhile, Neural Architecture Search (NAS) is becoming increasingly important as a technique to discover novel deep learning architectures. However, early NAS methods based on reinforcement learning or evolutionary algorithms suffered from ludicrous computational and memory costs. In contrast, differentiable NAS, a gradient-based approach, has the advantage of being much more efficient and has thus retained most of the attention in the past few years. In this article, we present NASiam, a novel approach that uses for the first time differentiable NAS to improve the multilayer perceptron projector and predictor (encoder/predictor pair) architectures inside siamese-networks-based contrastive learning frameworks (e.g., SimCLR, SimSiam, and MoCo) while preserving the simplicity of previous baselines. We crafted a search space designed explicitly for multilayer perceptrons, inside which we explored several alternatives to the standard ReLU activation function. We show that these new architectures allow ResNet backbone convolutional models to learn strong representations efficiently. NASiam reaches competitive performance in both small-scale (i.e., CIFAR-10/CIFAR-100) and large-scale (i.e., ImageNet) image classification datasets while costing only a few GPU hours. We discuss the composition of the NAS-discovered architectures and emit hypotheses on why they manage to prevent collapsing behavior.

Jigsaw-ViT: Learning jigsaw puzzles in vision transformer

The success of Vision Transformer (ViT) in various computer vision tasks has promoted the ever-increasing prevalence of this convolution-free network. The fact that ViT works on image patches makes it potentially relevant to the problem of jigsaw puzzle solving, which is a classical self-supervised task aiming at reordering shuffled sequential image patches back to their original form. Solving jigsaw puzzle has been demonstrated to be helpful for diverse tasks using Convolutional Neural Networks (CNNs), such as feature representation learning, domain generalization and fine-grained classification. In this paper, we explore solving jigsaw puzzle as a self-supervised auxiliary loss in ViT for image classification, named Jigsaw-ViT. We show two modifications that can make Jigsaw-ViT superior to standard ViT: discarding positional embeddings and masking patches randomly. Yet simple, we find that the proposed Jigsaw-ViT is able to improve on both generalization and robustness over the standard ViT, which is usually rather a trade-off. Numerical experiments verify that adding the jigsaw puzzle branch provides better generalization to ViT on large-scale image classification on ImageNet. Moreover, such auxiliary loss also improves robustness against noisy labels on Animal-10N, Food-101N, and Clothing1M, as well as adversarial examples. Our implementation is available at https://yingyichen-cyy.github.io/Jigsaw-ViT.

CGUN-2A: Deep Graph Convolutional Network via Contrastive Learning for Large-Scale Zero-Shot Image Classification

Taxonomy illustrates that natural creatures can be classified with a hierarchy. The connections between species are explicit and objective and can be organized into a knowledge graph (KG). It is a challenging task to mine features of known categories from KG and to reason on unknown categories. Graph Convolutional Network (GCN) has recently been viewed as a potential approach to zero-shot learning. GCN enables knowledge transfer by sharing the statistical strength of nodes in the graph. More layers of graph convolution are stacked in order to aggregate the hierarchical information in the KG. However, the Laplacian over-smoothing problem will be severe as the number of GCN layers deepens, which leads the features between nodes toward a tendency to be similar and degrade the performance of zero-shot image classification tasks. We consider two parts to mitigate the Laplacian over-smoothing problem, namely reducing the invalid node aggregation and improving the discriminability among nodes in the deep graph network. We propose a top-k graph pooling method based on the self-attention mechanism to control specific node aggregation, and we introduce a dual structural symmetric knowledge graph additionally to enhance the representation of nodes in the latent space. Finally, we apply these new concepts to the recently widely used contrastive learning framework and propose a novel Contrastive Graph U-Net with two Attention-based graph pooling (Att-gPool) layers, CGUN-2A, which explicitly alleviates the Laplacian over-smoothing problem. To evaluate the performance of the method on complex real-world scenes, we test it on the large-scale zero-shot image classification dataset. Extensive experiments show the positive effect of allowing nodes to perform specific aggregation, as well as homogeneous graph comparison, in our deep graph network. We show how it significantly boosts zero-shot image classification performance. The Hit@1 accuracy is 17.5% relatively higher than the baseline model on the ImageNet21K dataset.