Research On The Robustness Of CNN Features For Image Classification

Image classification is one of the fundamental tasks in the field of computer vision,with the goal of predicting the category of objects in an image.The performance of image classification tasks has a significant impact on various computer vision tasks such as object detection and image segmentation.In recent years,Convolutional Neural Networks(CNNs)have achieved breakthroughs in image classification tasks.However,factors like variations in viewpoint,scene changes,and Gaussian noise that are commonly present during the image acquisition process can affect the robustness of feature extraction by convolutional neural networks,leading to a decline in classification performance.Therefore,researching how to enhance the robustness of features extracted by neural networks is of great importance.This paper addresses three issues related to the robustness of feature:insufficient robustness to image rotation transformations,insufficient robustness to background variations,and the lack of robustness enhancement models applicable to various image transformations.To enhance the feature’s robustness to rotation transformations,the paper first introduces orientation adaptive convolutional neural networks(OACNN),to extract multi-directional information in an image.Secondly,to improve feature robustness to background variations,the paper proposes an Attention Weight Block(AWB),which designs Attention Weight Convolution layers to independently suppress background information for each convolutional kernel.Finally,to address the issue of existing robustness enhancement algorithms lacking mathematical models applicable to various image transformations,the paper establishes a Feature Robustness-enhancement Regularization Model(FRRM)and introduces robustness enhancement regularization terms.It mathematically models different feature robustness enhancement problems from a regularization perspective.The research contributions of this paper are as follows:To address the issue of inadequate robustness of features to image rotation transformations,current solutions primarily rely on data augmentation.However,features trained using data augmentation only robust to the angle transformations included in the augmentation and not to other angle variations.Furthermore,data augmentation leads to an increment in the number of training samples,resulting in an increment in computational power and time consumption.To mitigate the problem,this paper incorporates prior knowledge of image rotation transformations into the design of CNNs’ structure.It introduces orientation adaptive convolutional kernels(OA kernels)and orientation adaptive max pooling(OA max pooling),creating OACNN.OACNN can extract robust feature for image rotation transformations without data augmentation,thereby improving the accuracy of classifying test images with rotation transformations.Specifically,OA kernels can adaptively rotate convolutional kernels based on the orientation of input feature maps,extracting semantically relevant information from multiple directions and producing vector features that represent multi-directional information.OA max pooling spatially divides feature maps based on their orientation information,ensuring that the same contour has the same pooling region on different directional feature maps,enhancing the CNN’s ability to extract rotation-robust features.Experimental results demonstrate that compared to existing algorithms,OACNN extracts more robust features for image rotation transformations with fewer parameters.In particular,in the task of rotated image classification,OACNN achieves an accuracy on the CIFAR-10 dataset that is 5.97%higher than that of ORN.To address the issue of insufficient robustness of features to background variations,existing attention mechanisms enhance their robustness to background changes by assigning different attention coefficients to feature maps in spatial or channel dimensions.This mechanism assumes that the background information needed to be suppressed is the same for each convolutional kernel in the next convolutional layer.However,since each convolutional kernel extracts different features and focuses on different foreground information,the background information that needs to be suppressed is also different.Therefore,existing attention mechanisms struggle to finely suppress irrelevant background information based on the features extracted by each convolutional kernel.To alleviate this problem,this paper proposes the AWB to model the dependency between convolutional kernels and the content of the image.The convolutional kernel weights in AWB can automatically adjust itself based on the content of the input image during the test phase.This paper applies AWB to both the convolutional layer and batch normalization layer,and evaluates it on various datasets.Experimental results demonstrate that AWB enhances the anti-interference ability of various object detection and semantic segmentation algorithms to background information,thereby improving the classification and localization capabilities of these algorithms for foreground information.Additionally,in image classification tasks,AWB consistently outperforms existing attention mechanisms.Under similar parameter and time consumption conditions,AWB achieves an accuracy on the CIFAR-100 and Tiny-ImageNet datasets that is 1.0%and 1.1%higher,respectively,compared to the Efficient Channel Attention(ECA)mechanism.To address the issue of the lack of robustness enhancement models applicable to various image transformations in existing algorithms,this paper designs robustness enhancement regularization terms and introduces a FRRM.It transforms the feature robustness enhancement problems for different image transformations into a minimization problem of robustness enhancement regularization terms.The paper first defines a robustness measure for features at any sample point and further designs robustness enhancement regularization terms within the feature distribution region.Then,the effectiveness of these regularization terms is validated in robustness enhancement tasks for rotation transformations,background variations,and Gaussian noise.Specifically,despite the different motivations behind the introduction of OA kernels and attention mechanisms,this paper demonstrates that existing OA kernels and attention mechanisms implicitly optimize the robustness enhancement regularization terms by regularizing the CNN.This,in turn,enhances the network’s robustness to image rotation transformations and background variations.Furthermore,this paper explicitly incorporates robustness enhancement regularization terms for image rotation transformations,background variations,and Gaussian noise into the loss function and optimizes them.Experimental results show that explicitly optimizing the robustness enhancement regularization terms improves the robustness of image features to various image transformations,thereby simultaneously increasing the classification accuracy of the model for images with rotation transformations,background variations,and Gaussian noise.In summary,this paper investigates three issues related to the robustness of CNN features:insufficient robustness to image rotation transformations,insufficient robustness to image background variations,and the lack of robustness enhancement models applicable to various image transformations.To mitigate these problems and address the shortcomings of existing algorithms,this paper proposes the OACNN and the AWB to enhance feature robustness to image rotation transformations and background variations.Furthermore,the paper introduces a robustness enhancement model applicable to various image transformations,namely the FRRM.FRRM not only provides guidance for the design of feature robustness enhancement algorithms but also helps researchers better understand the principles behind existing feature robustness enhancement algorithms and their ability to extract robust features.