Research On Classification Algorithm Of Commodity Image Based On Depth Learning

As a new trade situation,E-commerce has brought infinite space for the development of online marketing.Online shopping mode has become a product of the new era of e-commerce.It is an inevitable requirement for the development of the Internet to realize the rapid classification of commodity images.The sale of commodities on the network mainly depends on the information transmission of commodity images.The application of convolution neural network has effectively changed the shortcomings of manual description of image information,such as content deviation,heavy workload and so on.It uses its unique convolution structure and deep network to learn the features of input samples,and realizes the automatic classification of network.In this paper,clothing Commodity images are selected for recognition and classification.Aiming at the problems of rich details,large amount of data and multi-angle deformation of the image itself,an improved network HSR-FCN is constructed by using deep learning algorithms such as full convolution network R-FCN,improved new Hyper features and spatial transformer network.The main work and innovations of this paper are as follows:(1)Aiming at the problem that the traditional convolution network has complex network structure and the classification effect is affected by object position transformation,a region-based full convolution network R-FCN is innovatively introduced.Firstly,the network locates the recognition target,and then classifies the objects in the positioning frame.Compared with the existing target recognition framework,R-FCN network has the fastest recognition speed and the highest recognition accuracy on VOC 2007 data set.Finally,the R-FCN network is constructed to classify the face-to-face and multi-angle deformed garment images respectively.The experimental results show that the recognition accuracy of the network is better,but the network training time is long,and the robustness of multi-angle deformed garment images is poor,and the recognition and classification accuracy is low.(2)Aiming at the problem that the traditional R-FCN network is weak in feature learning and takes a long time to train,a new feature fusion method is constructed,which divides the convolutional neural network of deep residual network into five layers and uses the middle layer as the basic network and the lowest pool as the largest pool.The high-leveldeconvolution method obtains the feature plane with the same resolution.After compression and fusion,a new Hyper feature is generated.The location information and semantic information of the image are taken into account.The feature learning level of the training process is enhanced,and the recognition accuracy of the garment image is guaranteed while the number of iterations is reduced.In addition,the convolution neural network is placed before the Ro I Pooling layer,which reduces the convolution of the last layer of the network,shortens the time spent in the network iteration and improves the timeliness.The experimental results show that the improved network HR-FCN improves the overall training time of the network,and further improves the recognition accuracy of the image,but the recognition of multi-angle deformed garment image is not good.(3)In order to improve the low recognition rate of multi-angle deformed garments,a new network HSR-FCN is constructed based on the improved HR-FCN network,which integrates two-layer space conversion network STN.The network transforms and aligns the input image and output feature map respectively in the training process.It improves the difficulty of feature extraction due to the garment angle problem and the poor effect of feature learning.The problem.Finally,the experimental samples are selected to compare the classification performance with the original network in the same experimental environment.The experimental results show that the improved network HSR-FCN constructed in this paper can significantly enhance the recognition of multi-angle deformed garment images without significantly increasing the training time.