Fine-Grained Classification Of Garment Images Based On Convolutional Neural Networks

At present,the image retrieval methods of shopping websites are mostly based on keywords.Businessmen label the name,size and price of product by keywords.This method of labeling is labor intensive and difficult to describe the full characteristics of the product.Although some scholars use deep learning networks to classify product images,the classification accuracy is not high due to the small samples and the poor performance of feature extraction.This paper presents a method of feature extraction and classification of product images based on convolutional neural networks.The main work is as follows:(1)Summarized the research on status and methods of domestic and international product images classification.Through reading a large number of paper on product images classification,this paper analyzed the problems and limitations of current image classification methods.Summarized and compared these methods,and finally determined the main research contents and methods of this paper.(2)Introduced the relevant basic theories of deep learning.The common models of deep learning are outlined,including deep self-encoding neural networks,deep belief networks,restricted Boltzmann machines,and convolutional neural networks.The related basic theories and common structural modules of convolutional neural networks,including convolutional layer,pooling layer,activation function,softmax classifier,normalization and dropout.This paper proposed the method of assessing network classification performance by calculating the convolution kernel capacity and coverage.Finally,introduced the deep learning framework Caffe,configured and run the Minist and Cifar-10 database.(3)This paper designs and trains a deep convolutional neural networks that the convolution kernel size and the order of network connection are based on the high efficiency of the filter capacity and coverage.The Caltech256 public image database and the homemade database were used data augmentation technology,and realized the garment imagesclassification under large samples.The classification accuracy of the deep convolutional neural network was 94.8% and 92.1%.Visualizing the features of each convolutional layer to obtain the feature extraction of the images and the corresponding mean and standard deviation are beneficial to debugging and optimizing the network structure and parameters.(4)This paper proposed a shallow convolutional neural network,and the validity and feasibility of the network were proved by calculating its convolution kernel capacity and coverage.The Image Net-1000 database was used for training,and then the network parameters were fine-tuned to obtain a pre-training model.This paper used the Faster-RCNN algorithm to detect image targets,eliminating the interference of complex backgrounds on network training.The experimental results showed that the classification accuracy of the shallow convolutional neural network reached 90.6%.It was similar to the deep networks classification accuracy.It solved the problem that the deep convolutional neural network had long training time and low resource utilization.