Research On Clothing Detection And Classification Based On Deep Convolutional Neural Network

In recent years,with the rapid development of artificial intelligence technology and smart devices,technologies such as clothing identification,clothing detection,clothing recommendation,and virtual dressing have become more and more popular.Since these technologies such as clothing recognition and detection are easily affected by the posture of the person,the size of the clothes,the background of the clothes,and the occlusion between the clothes,it is a challenge to explore an algorithm with high practicability,good robustness,and high accuracy.How to solve these difficulties in clothing,how to create a suitable clothing application model from machine learning has become a hot issue in machine learning,image processing and recognition,pattern recognition.This chapter has carried out research on clothing type identification,clothing attribute classification and clothing detection,and has completed the following work:1.A clothing detection based on soft non-maximum value-based multi-scale network is proposed.The algorithm has the problems that the detection frame is easy to overlap and the small clothing parts are difficult to detect.The multi-scale network extracts the trained ResNet features,inputs them into 5 RPN(Region Proposal Networks)networks and generates a large number of detection frames.Filter the suitable clothing frame with soft non-maximum value,then map the detection frame to the characteristics of ResNet,and finally use the multi-tasking classifier to identify the clothing category and return the frame.Experiments show that the proposed algorithm MAP reaches 95.2%,and the recall rate of 12 types has a good performance.2.A classification of clothing attributes based on the simple cross-layer inception-v4 algorithm is proposed.The algorithm introduces the DenseNet dense connection method on the simple cross-layer inception-v4 model for the multi-task problem of clothing category 26 clothing attributes,and optimizes the network parameters in the simple cross-layer inception-v4 network.In addition to reducing the amount of network computing,in addition,the cen function is introduced to adapt to the classification of clothing attributes.The proposed algorithm shares network features in a simple cross-layer inception-v4 network,and 26 classifiers are used to classify 26 attributes.The experimental results show that the average accuracy of the method and the accuracy of the 26 attributes are better than other literatures.3.A classification of clothing types based on 55-layer convolutional neural network under migration learning is proposed.The algorithm first performs migration learning on the 55 th layer of the network,then trains and extracts 55 layers of detailed features,and then increases the multi-layer linear inference to achieve higher clothing recognition accuracy.The main idea of the algorithm is to solve the complex problem of clothing characteristics by utilizing the excellent representation ability of the deep network,increasing the linear operation of the network and making full use of the 55-layer feature information.The experimental results show that the accuracy of the algorithm is significantly higher than that of other algorithms on the ACS data,and the network can learn more detailed clothing features.