| With the continuous development of e-commerce,the amount of clothing image data on the Internet has increased sharply.There are some problems such as unstable classification results,high labor cost and slow speed in the manual classification of clothing,which cannot meet the current practical needs.Among them,the image of women's wear is not only huge in number,but also a wide variety of categories,which makes classification more difficult.Therefore,the algorithm for image classification of women's wear is a meaningful research direction.In recent years,convolutional neural networks have made great breakthroughs in many fields of computer vision.In this paper,based on convolutional neural network,the classification algorithm of women's dress images is studied,and the following work is mainly completed:Firstly,the theoretical basis and basic concepts of convolution neural network are introduced.This paper analyzes the difficult problems existing in the application of image classification of women's wear at present,designs and improves the algorithm model pertinently,and constructs the structure of image classification algorithm based on neural network.There are a wide variety of images of women's wear,and the differences between different types of images are relatively small,most of which are only reflected in the details of some key parts.In addition,in the shooting of pictures,clothing often produces some distortion and deformation,and some areas are covered.Further increase the difficulty of women's image classification.In this paper,the structure of the traditional classification algorithm based on convolutional neural network is improved,and the feature information of key parts is utilized to better distinguish the similar categories.For this reason,this paper designs a method for locating the key points of clothing images,and optimizes the positioning method of key points through multiple experiments,so as to facilitate the extraction of feature information of key points.The final experimental results show that the model with key point feature information is more advantageous in accuracy and the robustness of the model is improved.In addition,in order to further improve the performance of the model,background information in the input image is removed to reduce interference.The preprocessing operation of input image is designed.The effect of pretreatment operation was compared by experiment.In practical applications,convolutional neural networks are limited by the level of hardware and the amount of training data.In this paper,the structure of the model is improved,and the calculation method of "block convolution" is proposed to replace the traditional convolution layer,which reduces the number of parameters and calculation of the model.As the number of parameters decreases,the complexity of the model decreases,and the risk of overfitting is reduced without increasing the amount of training data.At the same time,due to the reduction of computation,the training process of the model is accelerated,which is convenient for the debugging of superparameters.The prediction speed is also improved,which improves the practicability of the model.The experimental results show that the accuracy of the model is improved and the detection efficiency of the model is improved because the over-fitting phenomenon is inhibited.Finally,this paper describes the training process of the model in detail,including data set selection,data preprocessing and model fusion of the output results.Through the comparison of experimental data,the effect of the algorithm improvement is demonstrated.At the end of the paper,this paper summarizes the relevant work done in this paper,and points out some deficiencies in the current work and future improvement direction. |
