基于改进ShuffleNet v1的服装图像分类算法

Clothing image classification algorithm based on improved ShuffleNet v1

针对服装图像分类模型体积较大,缺少细分类的问题,提出基于改进ShuffleNet v1的服装图像分类算法。该算法以ShuffleNet v1为基础,通过优化模块的堆叠次数和网络层通道数来降低模型的计算量,满足算法的实时性要求;嵌入通道和空间注意力模块,使得模型关注重要的特征信息,抑制无用的特征信息;设计非对称多尺度特征融合模块,加强模型的特征提取能力。结果表明:所提算法在自建的衬衫服装数据集中准确率为88.31%,分别高于ShuffleNet v1、ShuffleNet v2、MobileNet v2和ResNet50模型2.77%、3.69%、1.98%、0.62%;所提算法在DeepFashion的部分数据集中也取得了不错的效果,验证了所提算法的有效性与通用性;与基础模型相比,所提模型的参数量仅为0.73M,模型参数量减少了约60%,实现了模型准确率和推理速度的提升。

In recent years, with the development of the Internet economy, compared with offline sales of clothing, online shopping has broken the time and geographical restrictions, and has gradually become one of the mainstream consumption methods with distinctive features such as various categories and affordable prices. According to the China E-Commerce Report 2021 released by the Ministry of Commerce of the People’s Republic of China, in 2021, the national online retail witnessed rapid growth, with the online retail volume reaching 13.09 trillion yuan, of which footwear and clothing products account for the largest proportion, reaching 22.94%. With the increasing demand for clothing, consumers’ requirements for clothing retrieval methods are also increasing. Consumers hope to retrieve more accurate results based on their own needs and styles. Therefore, the classification of clothing needs to be meticulous and accurate.To promote the accurate integration of clothing retrieval results of e-commerce platforms and consumer demand, it is necessary to further enrich the clothing retrieval methods of e-commerce platforms. Aiming at the problem of large volume and lack of fine classification of garment image classification models, a garment image classification algorithm based on improved ShuffleNet v1 is proposed. Based on ShuffleNet v1, the algorithm reduces the computational load of the model by optimizing the number of module stacks and network layer channels to meet the real-time requirements of the algorithm. Furthermore, the channel and spatial attention module is embedded to make the model focus on important feature information and suppress useless feature information. Finally, the asymmetric multi-scale feature fusion module is designed to enhance the feature extraction ability of the model. The results show that the accuracy of the proposed algorithm in the self-built shirt and clothing dataset is 88.31%, which is 2.77%, 3.69%, 1.98% and 0.62% higher than that of ShuffleNet v1, ShuffleNet v2, MobileNet v2 and ResNet50 models respectively. The proposed algorithm has also achieved good results in some datasets of DeepFashion, verifying the effectiveness and universality of the proposed algorithm. Compared with the basic model, the parameters of the proposed model are only 0.73M, and the parameters of the model are reduced by about 60%, which improves the accuracy and reasoning speed of the model.This paper proposes an algorithm to achieve high-precision and less time-consuming shirt clothing classification, which has practical application value for helping merchants to subdivide clothing in the early stage. At the same time, it has good academic significance and reference value for the research of similar fine classification and lightweight problems. However, due to the limitation of the dataset, only a few categories of shirts such as stripes, lattices and spots are supported. In real life, a certain type of clothing has different forms and more categories. How to collect effective data or expand datasets is a direction worth exploring in the future.