Research On Image Detection And Identification Of Clothing Based On Multi-weight Residual Network

At present,garment retrieval technology of most of the e-commerce platforms is still based on keyword search method,this technique often cannot give feedback to the user's favorite result,and content-based image retrieval method is also called to search by image,will bring great convenience to users' search clothing.There are a lot of problems that need to be solved in search by image: there are many kinds of clothes and the difference between many categories is very small.When using search by image,the background of the clothing picture that user enters is complex;Most of the clothing datasets are unbalanced in data distribution,and these problems greatly affect the training of clothing classification model.This thesis studies the classification and target detection of clothing images in unbalanced clothing datasets.The main work of this thesis is as follows:(1)In view of the problem of poor classification effect of the common shallow model VGG16 and AlexNet on multi-category datasets,this thesis builds a betterstructured residual network ResNet-50 P.This network consists of three small convolution kernel residual modules,and compared to the traditional two layers of residual unit module,the calculation amount is the same,but a deeper residual network can be constructed,and the learning ability of the model is enhanced.Then use the preactivation method in the residual unit module to further optimize the residual network and improve the accuracy of the model.(2)In view of the problem of large computation in the target detection model based on the regional proposal box,this thesis designed the Light Fashion R-CNN model based on the network structure of the target detection model R-FCN.In the Light Fashion R-CNN model,the shared convolutional layer used for feature extraction adopts the ResNet-50 P network structure.In the detection and identification module,the use of large convolution kernel separable parallel convolution layers is to improve the effectiveness of feature maps,at the same time,by reducing the number of positionsensitive score maps to reduce network computing.(3)In order to solve the problem of unbalanced amount of data between classes in DeepFashion,a multi-weight network is designed based on the network structure of ResNet-50 P to improve the classification accuracy of categories with fewer numbers.From the perspective of increasing the number of misclassified loss values in the categories with fewer numbers,two weight setting strategies are designed.(4)Based on the above three improvements,multiple sets of experiments were designed on the DeepFashion dataset.The applicability of the network models proposed in this thesis is analyzed from several perspectives.The experimental results show that the resnet-50 p model has good performance in classification accuracy,image processing speed and generalization ability.The Light Fashion R-CNN network has good performance in classification accuracy.The multi-weight network effectively improves the training effect of categories with fewer numbers on unbalanced datasets and improves the classification accuracy.This thesis study the method of image detection and identification of unbalance garment dataset,and it is suggested to replace the current shallow network with better network structure with better network structure.At the same time,the detection network module in the target detection model based on the regional proposal box is optimized.Finally,a multi-weight network is designed for the problem of data volume imbalance between data sets.The above work has certain reference significance to the research on the detection and identification of clothing images.