Clothing Classification Algorithm Based On Attention Mechanism And Feature Fusion

Deep learning-based techniques for clothing image classification have emerged as a prominent research area in the field of fashion.Given the rapid proliferation of clothing types and styles,accurate classification of apparel has become a crucial factor in online retail.In addition,improvements in the speed and accuracy of clothing classification can often lead to better results for other tasks as well.However,the complex backgrounds and diverse styles of clothing images make the features of clothing susceptible to interference during extraction.In addition,clothing classification that relies solely on texture characteristics may focus more on garment detailed information and ignore the overall shape information of clothing,resulting in low robustness and low accuracy.Consequently,this thesis introduces a clothing classification algorithm that leverages attention mechanisms and feature fusion techniques to enhance accuracy,which improves accuracy by extracting the shape and texture feature information of clothing separately.To address the above issues,the research work on the classification of clothing images in this thesis will consist of the following main aspects:(1)To enhance the extraction of bottom-level features while leveraging the full potential of high-level semantic features in clothing style classification,this thesis proposes an feature extraction algorithm,namely Pattern Net,which utilizes attention mechanisms and Depthwise Dilated Separable Convolution.This approach comprises two key modules: the Fused Inverted Residual(FIRConv)module and Depthwise Dilated Separable Attention Convolution(DDSAConv)module.The FIRConv module allows the underlying feature pattern and texture information of the garment image to be extracted sufficiently effectively through a fused inverted residual structure to facilitate semantic analysis of the features by the deeper network.Overlaying the use of the DDSAConv module in the deeper layers of the network with different dilated rates for convolution,allows the model to focus more on the global features of the garment by increasing the perceptual field of the model.According to the characteristics of the garment dataset,using different structures can make the model play a greater role in feature extraction at different stages and fully extract the garment texture information.(2)In order to effectively represent the feature information of garments,this thesis proposes a feature extraction algorithm based on shape features,which makes the network pay more attention to the shape information of garments while reducing the background interference of garment images.The method consists of two main components,the garment shape acquisition module and the shape classification network Shape Net,which first obtains the garment shape dataset through the garment shape acquisition module,and then acquires the garment shape features in Shape Net by using a larger convolutional kernel and an efficient attention mechanism.(3)In light of the difficulties associated with fine-grained classification of clothing,this thesis proposes a two-stream network structure PSClo Net to fully leverage shape features and multi-feature fusion to enable comprehensive utilization of clothing image features.Based on the differences in the way of feature extraction,the texture features of clothing images are effectively fused with shape features using the Feature Fusion Channel Enhancement(FFCE)module,thus significantly improving the accuracy of fashion image classification and achieving higher accuracy rates in diverse and variable styles,and the experimental results prove that the PSClo Net method is more effective than separate extracting a certain class of features.The experimental results demonstrate the superior performance of Pattern Net,as compared to mainstream classification networks,in the domain of apparel classification.Additionally,the fusion of Shape Net and Pattern Net within the PSClo Net network has been shown to further enhance the accuracy of clothing classification.These findings highlight the efficacy of the proposed methodology and underscore its potential to advance in the field of clothing image analysis.