| In view of the low utilization rate of shallow features extracted by convolutional neural network and the difficulty in utilizing the complementary advantages of high-level and low-level features,a method of fusing high-level and low-level features of convolutional neural network with selective feature connection mechanism(SFCM)is proposed to improve the accuracy of image recognition.Firstly,the low-level features of convolutional neural network are selected,and the dimension of low-level features is reduced by average pooling to output low-level features with more details;Secondly,according to the feature that the larger the value of high-level feature elements of convolutional neural network is,the more critical the corresponding position feature is,the importance score of element position of feature graph is calculated;Finally,the importance score is used to enhance or suppress the element information of low-level features to get the low-level features that are conducive to classification or recognition,and the low-level features are mapped to the level of high-level features to fuse the high-level and low-level features,so as to realize the complementary of high-level and low-level features of convolutional neural network,obtain the fusion features that are conducive to image recognition,and improve the image recognition rate of convolutional neural network.In the experiment,CIFAR-10 and CIFAR-100 datasets are selected for image recognition test.The results show that the convolution neural network based on selective feature connection has higher accuracy than the original model in image recognition on the self-defined shallow convolution neural network,Alex Net and VGG models.The image recognition rate is improved by 2.3%,1.8% and 2.1% respectively on CIFAR-10 datasets,In CIFAR-100 data set,it is improved by 4.8%,3% and 5.5% respectively,which proves that the method makes good use of high and low level image features and is conducive to image recognition. |
PDF Full Text Request