부산대학교 도서관

Clothing category classification is a key component of clothing analysis, and the classification performance largely depends on the method's ability to capture advanced clothing features, such as texture and contour information. In the frequency domain, different frequency bands of an image can represent texture and contour information at different levels, which cannot be well separated in the spatial domain. Therefore, feature extraction methods based on spatial domain information may lack the ability to analyze texture and contour information at different levels. This paper presents a new network architecture called FDENet, which introduces several innovative ideas to enhance performance. The main concepts of the framework are: 1) we utilize the spectrum enhancement module to obtain frequency band information from clothing images, which is used to represent the multilevel texture and contour information of the clothing, 2) we construct an information interaction module using depthwise separable convolution and residual blocks, which is utilized for the initial extraction and fusion of multi-level texture and contour features while reducing information redundancy, and 3) we propose a new clothing classification network based on the frequency domain enhancement approach, which can effectively extract advanced clothing features and achieve higher classification accuracy. Compared to the existing state-of-the-art classification models, our method achieves an improvement of 2.13% and 1.3% in Top-1 accuracy on the public datasets Clothing1M and Deepfashion, respectively.