Image Classification Datasets Research Articles

Reconstruction error based implicit regularization method and its engineering application to lung cancer diagnosis

The automatic diagnosis of lung cancer via artificial intelligence faces two hotspot issues: (1) insufficient data and (2) excessive redundant information, which make it difficult for convolutional neural networks (CNNs) to learn discriminative information of lung cancer. In this paper, we present the reconstruction error based implicit regularization method (REbIRM) that regularizes CNNs at the loss layer. During each training iteration, the reconstruction errors introduced by the two-stage discriminative auto-encoder are used to sharpen the generalization ability of deep CNNs by improving the decision boundary. In the application process, the trained deep CNN is used for completing computed tomography (CT) diagnostics. The main clinical benefit of our approach is that it is domain independent, requiring no specialized knowledge, and can therefore be applied to different types of datasets. To the best of our knowledge, this is the first attempt to implicitly regularize CNNs based on the reconstruction errors. Finally, experimental results on three CT image classification datasets show that REbIRM can achieve impressive results and that, in conjunction with Dropout, it obtains the state-of-the-art performance. REbIRM is also robust to the selection of hyper-parameters and only has the sublinear influence on the convergence of deep CNNs. Besides, empirical and theoretical evidence are provided to indicate that REbIRM prefers to converges in a constrained parameter space with flatter minima, which explains why it can generalize to new data. Finally, the nature of REbIRM is further explored through visualization techniques to analyze how it works in training deep CNNs.

Deep Recyclable Trash Sorting Using Integrated Parallel Attention.

Sorting recyclable trash is critical to reducing energy consumption and mitigating environmental pollution. Currently, trash sorting heavily relies on manpower. Computer vision technology enables automated trash sorting. However, existing trash image classification datasets contain a large number of images without backgrounds. Moreover, the models are vulnerable to background interference when categorizing images with complex backgrounds. In this work, we provide a recyclable trash dataset that supports model training and design a model specifically for trash sorting. Firstly, we introduce the TrashIVL dataset, an image dataset for recyclable trash sorting encompassing five classes (TrashIVL-5). All images are collected from public trash datasets, and the original images were captured by RGB imaging sensors, containing trash items with real-life backgrounds. To achieve refined recycling and improve sorting efficiency, the TrashIVL dataset can be further categorized into 12 classes (TrashIVL-12). Secondly, we propose the integrated parallel attention module (IPAM). Considering the susceptibility of sensor-based systems to background interference in real-world trash sorting scenarios, our IPAM is specifically designed to focus on the essential features of trash images from both channel and spatial perspectives. It can be inserted into convolutional neural networks (CNNs) as a plug-and-play module. We have constructed a recyclable trash sorting network building upon the IPAM, which produces an acuracy of 97.42% on TrashIVL-5 and 94.08% on TrashIVL-12. Our work is an effective attempt of computer vision in recyclable trash sorting. It makes a positive contribution to environmental protection and sustainable development.

Class-Level Logit Perturbation.

Features, logits, and labels are the three primary data when a sample passes through a deep neural network (DNN). Feature perturbation and label perturbation receive increasing attention in recent years. They have been proven to be useful in various deep learning approaches. For example, (adversarial) feature perturbation can improve the robustness or even generalization capability of learned models. However, limited studies have explicitly explored for the perturbation of logit vectors. This work discusses several existing methods related to class-level logit perturbation. A unified viewpoint between regular/irregular data augmentation and loss variations incurred by logit perturbation is established. A theoretical analysis is provided to illuminate why class-level logit perturbation is useful. Accordingly, new methodologies are proposed to explicitly learn to perturb logits for both the single-label and multilabel classification tasks. Meta-learning is also leveraged to determine the regular or irregular augmentation for each class. Extensive experiments on benchmark image classification datasets and their long-tail versions indicated the competitive performance of our learning method. As it only perturbs on logit, it can be used as a plug-in to fuse with any existing classification algorithms. All the codes are available at https://github.com/limengyang1992/lpl.

Optimal layer selection for latent data augmentation

While data augmentation (DA) is generally applied to input data, several studies have reported that applying DA to hidden layers in neural networks, i.e., feature augmentation, can improve performance. However, in previous studies, the layers to which DA is applied have not been carefully considered, often being applied randomly and uniformly or only to a specific layer, leaving room for arbitrariness. Thus, in this study, we investigated the trends of suitable layers for applying DA in various experimental configurations, e.g., training from scratch, transfer learning, various dataset settings, and different models. In addition, to adjust the suitable layers for DA automatically, we propose the adaptive layer selection (AdaLASE) method, which updates the ratio to perform DA for each layer based on the gradient descent method during training. The experimental results obtained on several image classification datasets indicate that the proposed AdaLASE method altered the ratio as expected and achieved high overall test accuracy.

Dynamic Routing and Knowledge Re-Learning for Data-Free Black-Box Attack.

Deep learning models have emerged as strong and efficient tools that can be applied to a broad spectrum of complex learning problems and many real-world applications. However, more and more works show that deep models are vulnerable to adversarial examples. Compared to vanilla attack settings, this paper advocates a more practical setting of data-free black-box attack, for which the attackers can completely not access the structures and parameters of the target model, as well as the intermediate features and any training data associated with the model. To tackle this task, previous methods generate transferable adversarial examples from a transparent substitute model to the target model. However, we found that these works have the limitations of taking static substitute model structure for different targets, only using hard synthesized examples once, and still relying on data statistics of the target model. This may potentially harm the performance of attacking the target model. To this end, we propose a novel Dynamic Routing and Knowledge Re-Learning framework (DraKe) to effectively learn a dynamic substitute model from the target model. Specifically, given synthesized training samples, a dynamic substitute structure learning strategy is proposed to adaptively generate optimal substitute model structure via a policy network according to different target models and tasks. To facilitate the substitute training, we present a graph-based structure information learning to capture the structural knowledge learned from the target model. For the inherent limitation that online data generation can only be learned once, a dynamic knowledge re-learning strategy is proposed to adjust the weights of optimization objectives and re-learn hard samples. Extensive experiments on four public image classification datasets and one face recognition benchmark are conducted to evaluate the efficacy of our Drake. We can obtain significant improvement compared with state-of-the-art competitors. More importantly, our DraKe consistently achieves attack superiority for different target models (e.g., residual networks, and vision transformers), showing great potential for complex real-world applications.

A Clustering Pruning Method Based on Multidimensional Channel Information

Pruning convolutional neural networks offers a promising solution to mitigate the computational complexity challenges encountered during application deployment. However, prevalent pruning techniques primarily concentrate on model parameters or feature mapping analysis to devise static pruning strategies, often overlooking the underlying feature extraction capacity of convolutional kernels. To address this, the study first quantitatively expresses the feature extraction capability of convolutional channels from three aspects: global features, distribution metrics, and directional metrics. It explores the multi-dimensional information of the channels, calculates the overall expectation, variance, and cosine distance from the unit vector as the quantitative results of the channels. Subsequently, a clustering algorithm is employed to categorize the multidimensional information. This approach ensures that convolutional channels grouped within each cluster possess similar feature extraction capabilities. An enhanced differential evolutionary algorithm is utilized to optimize the number of clustering centers across all convolutional layers, ensuring optimal grouping. The final step involves achieving channel sparsification through the calculation of crowding distances for each sample within its designated cluster. This preserves a diverse subset of channels that are critical for maintaining model accuracy. Extensive empirical evaluations conducted on three benchmark image classification datasets demonstrate the efficacy of this method. For instance, on the ImageNet dataset, the ResNet-50 model experiences a substantial reduction in FLOPs by 58.43% while incurring a minimal decrease in TOP-1 accuracy of only 1.15%.

On-Edge Deployment of Vision Transformers for Medical Diagnostics Using the Kvasir-Capsule Dataset

This paper aims to explore the possibility of utilizing vision transformers (ViTs) for on-edge medical diagnostics by experimenting with the Kvasir-Capsule image classification dataset, a large-scale image dataset of gastrointestinal diseases. Quantization techniques made available through TensorFlow Lite (TFLite), including post-training float-16 (F16) quantization and quantization-aware training (QAT), are applied to achieve reductions in model size, without compromising performance. The seven ViT models selected for this study are EfficientFormerV2S2, EfficientViT_B0, EfficientViT_M4, MobileViT_V2_050, MobileViT_V2_100, MobileViT_V2_175, and RepViT_M11. Three metrics are considered when analyzing a model: (i) F1-score, (ii) model size, and (iii) performance-to-size ratio, where performance is the F1-score and size is the model size in megabytes (MB). In terms of F1-score, we show that MobileViT_V2_175 with F16 quantization outperforms all other models with an F1-score of 0.9534. On the other hand, MobileViT_V2_050 trained using QAT was scaled down to a model size of 1.70 MB, making it the smallest model amongst the variations this paper examined. MobileViT_V2_050 also achieved the highest performance-to-size ratio of 41.25. Despite preferring smaller models for latency and memory concerns, medical diagnostics cannot afford poor-performing models. We conclude that MobileViT_V2_175 with F16 quantization is our best-performing model, with a small size of 27.47 MB, providing a benchmark for lightweight models on the Kvasir-Capsule dataset.

Early Breast Cancer Detection Using Artificial Intelligence Techniques Based on Advanced Image Processing Tools

The early detection of breast cancer is essential for improving treatment outcomes, and recent advancements in artificial intelligence (AI), combined with image processing techniques, have shown great potential in enhancing diagnostic accuracy. This study explores the effects of various image processing methods and AI models on the performance of early breast cancer diagnostic systems. By focusing on techniques such as Wiener filtering and total variation filtering, we aim to improve image quality and diagnostic precision. The novelty of this study lies in the comprehensive evaluation of these techniques across multiple medical imaging datasets, including a DCE-MRI dataset for breast-tumor image segmentation and classification (BreastDM) and the Breast Ultrasound Image (BUSI), Mammographic Image Analysis Society (MIAS), Breast Cancer Histopathological Image (BreakHis), and Digital Database for Screening Mammography (DDSM) datasets. The integration of advanced AI models, such as the vision transformer (ViT) and the U-KAN model—a U-Net structure combined with Kolmogorov–Arnold Networks (KANs)—is another key aspect, offering new insights into the efficacy of these approaches in different imaging contexts. Experiments revealed that Wiener filtering significantly improved image quality, achieving a peak signal-to-noise ratio (PSNR) of 23.06 dB and a structural similarity index measure (SSIM) of 0.79 using the BreastDM dataset and a PSNR of 20.09 dB with an SSIM of 0.35 using the BUSI dataset. When combined filtering techniques were applied, the results varied, with the MIAS dataset showing a decrease in SSIM and an increase in the mean squared error (MSE), while the BUSI dataset exhibited enhanced perceptual quality and structural preservation. The vision transformer (ViT) framework excelled in processing complex image data, particularly with the BreastDM and BUSI datasets. Notably, the Wiener filter using the BreastDM dataset resulted in an accuracy of 96.9% and a recall of 96.7%, while the combined filtering approach further enhanced these metrics to 99.3% accuracy and 98.3% recall. In the BUSI dataset, the Wiener filter achieved an accuracy of 98.0% and a specificity of 98.5%. Additionally, the U-KAN model demonstrated superior performance in breast cancer lesion segmentation, outperforming traditional models like U-Net and U-Net++ across datasets, with an accuracy of 93.3% and a sensitivity of 97.4% in the BUSI dataset. These findings highlight the importance of dataset-specific preprocessing techniques and the potential of advanced AI models like ViT and U-KAN to significantly improve the accuracy of early breast cancer diagnostics.

Innovation in public health surveillance for social distancing during the COVID-19 pandemic: A deep learning and object detection based novel approach.

The Corona Virus Disease (COVID-19) has a huge impact on all of humanity, and people's disregard for COVID-19 regulations has sped up the disease's spread. Our study uses a state-of-the-art object detection model like YOLOv4 (You Only Look Once, version 4), a very effective tool, on real-time 25fps, 1920 X 1080 video data streamed live by a camera-mounted Unmanned Aerial Vehicle (UAV) quad-copter to observe proper maintenance of social distance in an area of 35m range in this study. The model has demonstrated remarkable efficacy in identifying and quantifying instances of social distancing, with an accuracy of 82% and little latency. It has been able to work efficiently with real-time streaming at 25-30 ms. Our model is based on CSPDarkNet-53, which was trained on the MS COCO dataset for image classification. It includes additional layers to capture feature maps from different phases. Additionally, the model's neck is made up of PANet, which is used to aggregate the parameters from various CSPDarkNet-53 layers. The CSPDarkNet-53's 53 convolutional layers are followed by 53 more layers in the model head, for a total of 106 completely convolutional layers in the design. This architecture is further integrated with YOLOv3, resulting in the YOLOv4 model, which will be used by our detection model. Furthermore, to differentiate humans The aforementioned method was used to evaluate drone footage and count social distance violations in real time. Our findings show that our model was reliable and successful at detecting social distance violations in real-time with an average accuracy of 82%.

Generation of Rule-Based Explanations of CNN Classifiers Using Regional Features

Although Deep Learning networks generally outperform traditional machine learning approaches based on tailored features, they often lack explainability. To address this issue, numerous methods have been proposed, particularly for image-related tasks such as image classification or object segmentation. These methods generate a heatmap that visually explains the classification problem by identifying the most important regions for the classifier. However, these explanations remain purely visual. To overcome this limitation, we introduce a novel CNN explainability method that identifies the most relevant regions in an image and generates a decision tree based on meaningful regional features, providing a rule-based explanation of the classification model. We evaluated the proposed method on a synthetic blob’s dataset and subsequently applied it to two cell image classification datasets with healthy and pathological patterns.

Deepdive: Leveraging Pre-trained Deep Learning for Deep-Sea ROV Biota Identification in the Great Barrier Reef

Understanding and preserving the deep sea ecosystems is paramount for marine conservation efforts. Automated object (deep-sea biota) classification can enable the creation of detailed habitat maps that not only aid in biodiversity assessments but also provide essential data to evaluate ecosystem health and resilience. Having a significant source of labelled data helps prevent overfitting and enables training deep learning models with numerous parameters. In this paper, we contribute to the establishment of a significant deep-sea remotely operated vehicle (ROV) image classification dataset with 3994 images featuring deep-sea biota belonging to 33 classes. We manually label the images through rigorous quality control with human-in-the-loop image labelling. Leveraging data from ROV equipped with advanced imaging systems, our study provides results using novel deep-learning models for image classification. We use deep learning models including ResNet, DenseNet, Inception, and Inception-ResNet to benchmark the dataset that features class imbalance with many classes. Our results show that the Inception-ResNet model provides a mean classification accuracy of 65%, with AUC scores exceeding 0.8 for each class.

Bi-Directional Ensemble Feature Reconstruction Network for Few-Shot Fine-Grained Classification.

The main challenge for fine-grained few-shot image classification is to learn feature representations with higher inter-class and lower intra-class variations, with a mere few labelled samples. Conventional few-shot learning methods however cannot be naively adopted for this fine-grained setting - a quick pilot study reveals that they in fact push for the opposite (i.e., lower inter-class variations and higher intra-class variations). To alleviate this problem, prior works predominately use a support set to reconstruct the query image and then utilize metric learning to determine its category. Upon careful inspection, we further reveal that such unidirectional reconstruction methods only help to increase inter-class variations and are not effective in tackling intra-class variations. In this paper, we introduce a bi-reconstruction mechanism that can simultaneously accommodate for inter-class and intra-class variations. In addition to using the support set to reconstruct the query set for increasing inter-class variations, we further use the query set to reconstruct the support set for reducing intra-class variations. This design effectively helps the model to explore more subtle and discriminative features which is key for the fine-grained problem in hand. Furthermore, we also construct a self-reconstruction module to work alongside the bi-directional module to make the features even more discriminative. We introduce the snapshot ensemble method in the episodic learning strategy - a simple trick to further improve model performance without increasing training costs. Experimental results on three widely used fine-grained image classification datasets, as well as general and cross-domain few-shot image datasets, consistently show considerable improvements compared with other methods.

GAP: A group-based automatic pruning algorithm via convolution kernel fusion

In recent years, the deployment and operation of convolution neural networks on edge devices with limited computing capabilities have become increasingly challenging due to the large network structure and computational cost. Currently, the mainstream structured pruning algorithms mainly compress the network at the filter or layer level. However, these methods introduce too much human intervention with large granularities, which may lead to unpredictable performance after compression. In this paper, we propose a group-based automatic pruning algorithm(GAP) via kernel fusion to automatically search for the optimal pruning structure in a more fine-grained manner. Specifically, we first adopt a novel nonlinear dimensionality reduction clustering algorithm to divide the filters of each convolution layer into groups of equal size. Afterwards, we encode the mutual distribution similarity of the kernels within each group, and its KL divergence is employed as an importance indicator to determine the retained kernel groups through weighted fusion. Subsequently, we introduce an intelligent searching module that automatically explore and optimize the pruned structure of each layer. Finally, the pruned filters are permutated to form a dense group convolution and fine-tuned. Sufficient experiments show that, on two image classification datasets, for five advanced CNN models, our GAP algorithm outperforms most extant SOTA schemes, reduces artificial intervention, and enables efficient end-to-end training of compact models.

Multi-source fully test-time adaptation

Deep neural networks have significantly advanced various fields. However, these models often encounter difficulties in achieving effective generalization when the distribution of test samples varies from that of the training samples. Recently, some fully test-time adaptation methods have been proposed to adapt the trained model with the unlabeled test samples before prediction to enhance the test performance. Despite achieving remarkable results, these methods only involve one trained model, which could only provide certain side information for the test samples. In real-world scenarios, there could be multiple available trained models that are beneficial to the test samples and are complementary to each other. Consequently, to better utilize these trained models, in this paper, we propose the problem of multi-source fully test-time adaptation to adapt multiple trained models to the test samples. To address this problem, we introduce a simple yet effective method utilizing a weighted aggregation scheme and introduce two unsupervised losses. The former could adaptively assign a higher weight to a more relevant model, while the latter could jointly adapt models with online unlabeled samples. Extensive experiments on three image classification datasets show that the proposed method achieves better results than baseline methods, demonstrating the superiority in adapting to multiple models.

Difficulty level-based knowledge distillation

Knowledge distillation (KD) enables a simple model (student model) to perform as a complex model (teacher model) by distilling the knowledge from a pre-trained teacher model. Existing soft-label distillation methods often use a fixed temperature value in the softmax function to prevent overconfidence in the distillation process. However, this approach can lead to the suppression of important ‘dark knowledge’ for non-target classes in difficult samples, while also over-smoothing the confidence values for easier samples. To address this issue, we propose a novel approach called difficulty level-based knowledge distillation (DLKD), which considers the difficulty level of each sample to distill refined knowledge with high or low confidence, depending on the sample’s complexity. Our method calculates the difficulty level based on the Euclidean distance between the teacher model’s predictions and the pruned teacher model’s predictions. Experimental results demonstrate that our DLKD method outperforms state-of-the-art methods on challenging samples, including those with noisy labels or augmented data, achieving superior results on CIFAR-100, FGVR, and ImageNet datasets for image classification.

A diffusion model multi-scale feature fusion network for imbalanced medical image classification research

Background and objectiveMedicine image classification are important methods of traditional medical image analysis, but the trainable data in medical image classification is highly imbalanced and the accuracy of medical image classification models is low. In view of the above two common problems in medical image classification. This study aims to: (i) effectively solve the problem of poor training effect caused by the imbalance of class imbalanced data sets. (ii) propose a network framework suitable for improving medical image classification results, which needs to be superior to existing methods. MethodsIn this paper, we put in the diffusion model multi-scale feature fusion network (DMSFF), which mainly uses the diffusion generation model to overcome imbalanced classes (DMOIC) on highly imbalanced medical image datasets. At the same time, it is processed according to the cropped image augmentation strategy through cropping (IASTC). Based on this, we use the new dataset to design a multi-scale feature fusion network (MSFF) that can fully utilize multiple hierarchical features. The DMSFF network can effectively solve the problems of small and imbalanced samples and low accuracy in medical image classification. ResultsWe evaluated the performance of the DMSFF network on highly imbalanced medical image classification datasets APTOS2019 and ISIC2018. Compared with other classification models, our proposed DMSFF network achieved significant improvements in classification accuracy and F1 score on two datasets, reaching 0.872, 0.731, and 0.906, 0.836, respectively. ConclusionsOur newly proposed DMSFF architecture outperforms existing methods on two datasets, and verifies the effectiveness of generative model inverse balance for imbalance class datasets and feature enhancement by multi-scale feature fusion. Further, the method can be applied to other class imbalanced data sets where the results will be improved.

OCIE: Augmenting model interpretability via Deconfounded Explanation-Guided Learning

Deep neural networks (DNNs) often encounter significant challenges related to opacity, inherent biases, and shortcut learning, which undermine their practical reliability. In this study, we address these issues by constructing a causal graph to model the unbiased learning process of DNNs. This model reveals that recurrent background information in training samples acts as a confounder, leading to spurious correlations between model inputs and outputs, causing biased predictions. To mitigate these problems and promote unbiased feature learning, we propose the Object-guided Consistency Interpretation Enhancement (OCIE) methodology. OCIE enhances DNN interpretability by integrating explicit objects and explanations into the model’s learning process. Initially, OCIE employs a graph-based algorithm to identify explicit objects within self-supervised vision transformer-learned features. Subsequently, it constructs class prototypes to eliminate invalid detected objects. Finally, OCIE aligns explanations with explicit objects, directing the model’s attention towards the most distinctive classification features rather than irrelevant backgrounds. Extensive experiments on different image classification datasets, including general (ImageNet), fine-grained (Stanford Cars and CUB-200), and medical (HAM) datasets, using two prevailing network architectures, demonstrate that OCIE significantly enhances explanation consistency across all datasets. Furthermore, OCIE proves particularly advantageous for fine-grained classification, especially in few-shot scenarios, by improving both interpretability and classification performance. Additionally, our findings highlight the impact of centralized explanations on the sufficiency of model decisions, suggesting that focusing explanations on explicit objects improves the reliability of DNN predictions. Our code is available at: https://github.com/DLAIResearch/OCIE.

Research on sports image classification method based on SE-RES-CNN model

As computer image processing and digital technologies advance, creating an efficient method for classifying sports images is crucial for the rapid retrieval and management of large image datasets. Traditional manual methods for classifying sports images are impractical for large-scale data and often inaccurate when distinguishing similar images. This paper introduces an SE module that adaptively adjusts the weights of input feature mapping channels, and a Res module that excels in deep feature extraction, preventing gradient vanishing, multi-scale processing, and enhancing generalization in image recognition. Through extensive experimentation on network structure adjustments, the SE-RES-CNN neural network model is applied to sports image classification. The model is trained on a sports image classification dataset from Kaggle, alongside VGG-16 and ResNet50 models. Training results show that the proposed SE-RES-CNN model improves classification accuracy by approximately 5% compared to VGG-16 and ResNet50 models. Testing revealed that the SE-RES-CNN model classifies 100 out of 500 sports images in 6 s, achieving an accuracy rate of up to 98% and a single prediction time of 0.012 s. This validates the model's accuracy and effectiveness, significantly enhancing sports image retrieval and classification efficiency. This validates the model's accuracy and effectiveness, significantly enhancing sports image retrieval and classification efficiency.

Neural Network Layer Algebra: A Framework to Measure Capacity and Compression in Deep Learning.

We present a new framework to measure the intrinsic properties of (deep) neural networks. While we focus on convolutional networks, our framework can be extrapolated to any network architecture. In particular, we evaluate two network properties, namely, capacity, which is related to expressivity, and compression, which is related to learnability. Both these properties depend only on the network structure and are independent of the network parameters. To this end, we propose two metrics: the first one, called layer complexity, captures the architectural complexity of any network layer; and, the second one, called layer intrinsic power, encodes how data are compressed along the network. The metrics are based on the concept of layer algebra, which is also introduced in this article. This concept is based on the idea that the global properties depend on the network topology, and the leaf nodes of any neural network can be approximated using local transfer functions, thereby allowing a simple computation of the global metrics. We show that our global complexity metric can be calculated and represented more conveniently than the widely used Vapnik-Chervonenkis (VC) dimension. We also compare the properties of various state-of-the-art architectures using our metrics and use the properties to analyze their accuracy on benchmark image classification datasets.

FPWT: Filter pruning via wavelet transform for CNNs

The enormous data and computational resources required by Convolutional Neural Networks (CNNs) hinder the practical application on mobile devices. To solve this restrictive problem, filter pruning has become one of the practical approaches. At present, most existing pruning methods are currently developed and practiced with respect to the spatial domain, which ignores the potential interconnections in the model structure and the decentralized distribution of image energy in the spatial domain. The image frequency domain transform method can remove the correlation between image pixels and concentrate the image energy distribution, which results in lossy compression of images. In this study, we find that the frequency domain transform method is also applicable to the feature maps of CNNs. The filter pruning via wavelet transform (WT) is proposed in this paper (FPWT), which combines the frequency domain information of WT with the output feature map to more obviously find the correlation between feature maps and make the energy into a relatively concentrated distribution in the frequency domain. Moreover, the importance score of each feature map is calculated by the cosine similarity and the energy-weighted coefficients of the high and low frequency components, and prune the filter based on its importance score. Experiments on two image classification datasets validate the effectiveness of FPWT. For ResNet-110 on CIFAR-10, FPWT reduces FLOPs and parameters by more than 60.0 % with 0.53 % accuracy improvement. For ResNet-50 on ImageNet, FPWT reduces FLOPs by 53.8 % and removes parameters by 49.7 % with only 0.97 % loss of Top-1 accuracy.