基于改进ResNet网络和迁移学习的服装图像风格识别研究OA

Traditional clothing image style recognition methods mainly rely on the successful extraction of ef-fective features,and these methods not only consume a lot of time and energy when processing images,but also have low recognition accuracy.In order to improve the performance of clothing image style recognition,this pa-per proposes a clothing image style recognition method based on the improved ResNet152 network and transfer learning.Firstly,the 7×7 convolutional kernel in the first layer structure of ResNet152 network is replaced by three 3×3 convolutional kernel combination layers,and secondly,the combination of"convolutional layer(Conv)+ batch normalization layer(BN)+ nonlinear activation function layer(Relu)"in the original residual unit is replaced by"batch normalization layer(BN)+ nonlinear activation function layer(Relu)+ convolutional layer(Conv)".These two improved methods effectively enhance the network performance and enable it to better capture clothing style features at different scales.The parameters of the ResNet152 network model trained on the ImageNet dataset are then migrated to the improved network,based on which the girl's clothing dataset is in-put to the network for training and validation as well as fine-tuning the network parameters.The results show that the proposed method in this paper has good training effect,and the recognition accuracy and convergence speed are better than other type recognition networks,which can better accomplish the task of girls'clothing style recognition.