Research And Application Of Hat Multi-classification Method Based On Convolution Neural Network

Faced with massive pictures,how to search and classify based on the content of pictures intelligently is a classic problem in the field of computer vision.Image classification is to distinguish whether an image contains the object or not.Content-based image classification technology is mainly divided into three steps: feature extraction of images,feature expression of extracted features and classification using classifiers.In order to solve the problem that it is difficult to search and classify massive images intelligently,this paper mainly studies the problem of multi-classification of hats,which is used for intelligent search and classification in specific situations.Firstly,classical image classification algorithms based on image content are studied and analyzed,which are feature extraction algorithm and classification algorithm,and KNN classification algorithm is implemented.The experimental results show that the accuracy and processing time of traditional image classification methods can not meet people's needs.In addition,the false detection rate for complex scenes and poor image quality is particularly high.After that,this paper mainly studies the multi-classification of hats based on convolution neural network.The main contents of this paper are as follows:(1)To solve the problem of multi-classification of hats,a five-layer convolution neural network structure is studied and designed.When designing the network structure,BN layers are added,then the convolution layer and the BN layer are merged together.The purpose of adding BN layer is to normalize the data,effectively solve the problem of gradient disappearance and gradient explosion,and accelerate the convergence of network and control over-fitting.The purpose of merging convolution layer with BN layer is to reduce computation,occupy more memory or display space,and improve the speed of forward inference.In addition,considering the computational complexity of network structure,the size of convolution cores in all convolution layers is 3x3,which reduces the computational complexity of network structure.(2)In view of the insufficient cap data set,the data augmentation function is added to the caffe framework.The caffe source code is modified to increase the function of data augmentation,and then the system performs image augmentation by assigning parameters to the network structure.The principle of the method is to increase the sample data by randomly rotating the image,cutting the image,changing the color of the image,changing the image saturation,changing the brightness of the image,changing the size of the image and so on.(3)In view of the poor image quality,the illumination pre-treatment is carried out.To a certain extent,the uneven illumination of the image changes the original appearance of the image,which makes the model test results wrong.In order to improve the detection effect of the model.The illumination pre-treatment of the image before the model test is studied to improve the model correction rate.(4)The optimized model is applied to the cap multi-classification intelligent search system based on image content.A simple web intelligent search system is designed and implemented to verify the results of model optimization.According to the network structure designed in this paper and the model trained by optimization method,the average positive detection rate can reach more than 0.96,which can be well applied to content-based intelligent search and classification scenes.