摘要
针对现有服装图像分类算法参数量较多、识别精度低的问题,提出了一种基于注意力机制和迁移学习的改进型MobileNet v2算法.首先,选取MobileNet v2作为特征提取网络,确保服装分类算法的整体轻量性.其次,将通道与空间注意力机制嵌入特征提取单元,自适应地选择和强化有用的特征信息,从而提高服装图像分类算法的识别精度.最后,通过迁移学习方法对模型进行参数初始化,使得模型能够从源域中获得先验知识.在Fashion MNIST数据集上的实验结果表明:所提算法的分类精度为93.28％,相较于ResNet50、EfficientNet v2_1、ShuffleNet v2 和 MobileNet v2模型,分别提高了1.85％、1.34％、3.86％和 3.17％;在DeepFashion数据集上的准确率为88.24％.此外,该算法参数量低至2.35M,单张图像推理速度仅为7.5 ms,在参数量基本不变的的情况下提升了分类精度与推理速度.
Abstract
With the continuous development of Internet technology,online clothing shopping has become one of the mainstream ways for people to shop.Consumers can easily browse and purchase various types of clothing from e-commerce platforms at home,without the need to visit physical stores in person.At the same time,online shopping platforms also offer more choices and competitive prices,which is one of the main reasons why people choose to shop for clothing online.According to market research data,as of 2021,the consumption scale of online clothing shopping in China had reached 3.7 trillion yuan,accounting for nearly half of the entire online shopping market,and this number is still growing.However,quickly and accurately classifying a large number of clothing images is a challenging task.Traditional clothing classification requires a lot of time and labor to classify,sort,and label,consuming a lot of manpower and time costs.Moreover,with the annual increase in the volume of online shopping orders,traditional manual classification methods are difficult to handle large amounts of data and cannot meet the needs of rapid classification and processing.On the contrary,clothing classification methods based on deep learning can learn more features and patterns through a large amount of data and iterative training,achieving higher accuracy without human intervention.Therefore,this article proposes a clothing classification algorithm based on improved MobileNet v2.The improved algorithm mainly embeds channel and spatial attention mechanisms into the basic unit of MobileNet v2 to form an attention mechanism basic unit,giving useful features greater weight,suppressing useless features,and enhancing the network's feature extraction ability.In addition,transfer learning is used to optimize the model parameters,and enhance the model's generalization ability and stability,so as to further improve the classification accuracy.Experimental results on the Fashion MNIST dataset show that the improved algorithm achieves an average accuracy rate of 93.28％,which is respectively 1.85％,1.34％,3.86％,and 3.17％higher than that of the ResNet50,EfficientNet v2_l,ShuffleNet v2,and MobileNet v2 models,effectively improving the problem of low accuracy in clothing image classification.The algorithm proposed in this article achieves high-precision clothing classification,which not only provides consumers with a better shopping experience but also provides e-commerce platforms with more accurate clothing recommendation services,and has high application value.However,the algorithm in this study focuses on single-label clothing image classification,and there are some limitations in recognizing multiple objects within a single image.Future research will emphasize expanding the clothing image classification task to multi-label classification,and further exploring how to handle multiple clothing categories in the same image,in order to apply it to a broader range of clothing classification scenarios.
维普
万方数据