基于改进ResNet50和迁移学习的服饰分类识别

Research on apparel classification recognition based on improved ResNet50 and transfer learning

传统的服饰分类方法主要依靠提取图像的纹理、颜色、边缘等特征,过程繁琐且分类精度不高。为了提高服饰图像分类性能,提出了一种基于改进ResNet50和迁移学习的服饰分类识别方法。首先,在ResNet50网络的STAGE5中引入两个池化核大小为2×2,步长为2的平均池化层并配合卷积核大小为1×1,步长为1的卷积层。其次,在最后一个卷积层后面融合卷积注意力机制模块(Convolutional Block Attention Module,CBAM)。这两个改进方法在降低特征图空间尺寸的同时保留了更多的信息,提升了模型的性能。最后,采用迁移学习方法,将ImageNet数据集上预训练权重迁移到改进的网络中,并使用服饰图像数据集对网络进行微调及验证。结果表明:改进后的ResNet50网络分类准确率可达90%,在Top1、Top3和Top5分类准确率上分别比原始的ResNet50提升了2.5%、0.4%、0.1%。此外,相较于现有的四个经典的卷积神经网络(GoogleNet,VGG-16,MobileNet_v2,AlexNet)有着更高的准确率,验证了其在服饰图像分类识别领域的优越性。

Traditional apparel classification methods mainly rely on extracting features such as texture,color,and edge of the image,which is a cumbersome process and has low classification accuracy.In order to improve the performance of apparel category classification,an apparel classification and recognition method based on improved ResNet50 and migration learning is proposed.Firstly,two average pooling layers with pooling kernel size of 2×2 and step size of 2 are added to STAGE5 of ResNet50 network together with a convolutional layer with convolutional kernel size of 1×1 and step size of 1.Secondly,the Convolutional Block Attention Module(CBAM)is fused behind the last convolutional layer.these two improvement methods make it possible to reduce the dimension of the feature map while retaining more information,which improves the performance of the model;lastly,a migration learning method is used to migrate the trained weights on the ImageNet dataset to the improved network,and the network is fine-tuned and validated using the dress image dataset.The results show that the accuracy of the improved ResNet50 network is up to 90%,which is 2.5%,0.4%,and 0.1%higher than the original ResNet50 in Top1,Top3,and Top5 classification accuracy,respectively.Meanwhile,it has higher accuracy than the existing four classical convolutional neural networks(GoogleNet,VGG-16,MobileNet_v2,AlexNet),which verifies the superiority of this model in the field of apparel image classification and recognition.