关键词: 服装图像, 属性预测, 注意力机制, ResNet50, 深度学习

Abstract: In recent years, with the popularity of online shopping, a large number of clothing images have emerged on the Internet. How to automatically extract key information from these massive clothing images has become a hot topic in current research. Through analyzing and identifying the relevant attributes of these clothing images and combining them with information such as price, sales volume and user comments, intelligent recommendations and trend predictions can be further achieved. This not only helps merchants grasp market demand in advance and formulate more accurate marketing strategies and business decisions but also provides designers with valuable creative inspiration. However, labeling the attributes of a large number of clothing images is also a tedious and costly task for online clothing sellers. Therefore, researching the classification and prediction of clothing image attributes has important practical significance and application value. 
To improve the prediction accuracy of clothing image attributes and to address the inefficiency of manual labeling of clothing image attributes, this paper proposes a clothing image attribute prediction method integrating the attention mechanism and improved ResNet50. This method improves the network structure of the ResNet50 model to adapt to the clothing multi-attribute prediction task and introduces the attention mechanism into the improved ResNet50 model to capture the detailed features of clothing attributes to improve the prediction accuracy. The method not only applies the improved deep learning algorithm to clothing attribute prediction, but also verifies the effectiveness of the method in clothing attribute prediction. It can effectively improve the accuracy of clothing image attribute prediction and identify attribute categories with superior prediction outcomes, providing new ideas for realizing the automatic labeling of clothing image attributes. The experimental results show that in the absence of the attention mechanism, the method based on the improved ResNet50 outperforms the traditional multi-label classification method in terms of accuracy, precision, recall, and F1 score, with the accuracy increasing by 25.96%. On the whole, the ResNet50 model performs better than the ResNet34, EfficientNet_V2, and VGG16 models in terms of accuracy, precision, recall, and F1 score. Compared with the method without the introduction of the CBAM attention mechanism, the ResNet50 method enhanced with CBAM improves the accuracy by 1.72%. In the prediction of each attribute category, the pattern, sleeve type, and style performed well, while the accuracy of the collar type is only 0.684, which is not good. In addition, the fabric and tightness show high accuracy but low recall. In future research, higher quality datasets can be built for training, and certain clothing attribute categories can also be studied separately to improve the prediction accuracy of the model. 

Key words: clothing images, attribute prediction, attention mechanism, ResNet50, deep learning