Research On The Algorithm Via Fusing The CNN Mid Layers Feature

In recent years,Convolution neural network(CNN)is a pattern recognition method combining artificial neural network and depth learning theory,and it has made great progress in image classification.Its weight sharing network structure makes it more similar to biological neural network,reducing the complexity of the network model,and greatly reducing the training network parameters.Unlike traditional image classification,convolution neural networks can automatically extract image features from a large number of data and classify them.However,how to further improve the performance of the algorithm is still one of the hot issues of academic research.Recent related studies have also shown that mid layers feature in CNN model can complement high-level semantic features missing and the defects of single feature.In this thesis mainly improves the accuracy of CNN recognition by means of feature fusion.The main idea of this algorithm is to extract the visual features of the middle layer by using the trained CNN model(AlexNet)on the caffe platform.Then,one-against-one algorithm is used to establish three layers(con5,fc6,fc7)SVM classifier,and finally we use the feature weight learning method to adjust the adaptive each feature learning coefficient.For each test image,three SVM classifiers are used to derive the decision values of each category.Then,three different weights are used to get the confidence value of the image,and the minimum error is adopted.The confidence value is the corresponding image type.The main innovations and achievements of this thesis include the following:1.Mid layers feature in CNN model can complement high-level semantic features missing and the defects of single feature.This thesis presents a fusion algorithm based on CNN hidden layer feature.2.When we use the SVM-based serial fusion mode,there is a drawback:it is not possible to flexibly adjust the weight of different layers of visual features to the classification,which has a significant effect on the output decision of the classifier.The parallel fusion mode uses the feature weight learning method to adjust the adaptive each feature learning coefficient,so we use the parallel fusion method based on SVM in this thesis.3.This thesis explores the problem of CNN hidden layer feature fusion image classification.Finally,it is proved that the effect of the classifier is better than that of other layers.4.Experiments were performed on the Caltech 256 database,and the classification effect of the combined classifier was increased to 95.83%.The features extracted from the hidden layer of CNN are suitable for SVM-based the combined classifier.