Research On Fine-Grained Image Recognition Based On Depth Separable Convolution

Fine-grained image recognition is a basic content in the research of computer vision and plays an indispensable role in it.At the same time,it has a deeply research in academic and practical applications.In recent years,thanks to the quickly growth and perfection of technology such as deep learning,the recognition technology of ordinary images has reached an upper level,while the task of fine-grained image recognition and classification still has a long way to go.The difficulty of fine-grained image classification is that there are some different types of images with very small differences,while the gap between images of the same type is very large.In addition,fine-grained image classification still has some problems that the main body of the image is not obvious,the background information is too complicated,and the number of data sets is not rich enough.Therefore,accurately extracting the features of the main body of the image has become an important breakthrough in solving fine-grained classification tasks.The traditional convolutional neural network can effectively complete the recognition of ordinary images,but the recognition accuracy of fine-grained images is not very high.Therefore,this paper proposes some new methods to experiment and verify the datasets commonly used in fine-grained image recognition tasks.The research work of this article includes:1.The dataset used in the experiment was analyzed and sorted out,and the relevant information of the dataset is introduced closely.2.VGG19,Resnet50 residual network and Inception V3 network were applied to carry out superficial identification research on the datasets.3.An attention mechanism has been introduced to achieve effective enhancement of fine-grained images through attention-guided data enhancement.Network structure combines attention mechanism and Inception V3 network(first use attention mechanism for data enhancement,then use attention pool to determine characteristics).Further identification studies of the dataset described in the article,and the recognition rate is further improved compared to the results of the first three methods of the experiment obtained from.4.After the above research,new network structure Xception network(fewer parameters and more accurate feature extraction structure)was introduced.Meanwhile,the method of feature extraction was improved by using depth separable convolution to locate the main body of images more accurately and extract the main features of images more accurately.Finally,in this article,we will experiment through some commonly used fine-grained image datasets to test the effectiveness of the methods in this article.At the same time,we comprehensively analyze the test results of various neural networks involved in the experiment.The classification accuracy of this method on AIR,CAR,CUB,DOG,and NABirds datasets reached 93.04%,94.52%,89.39%,91.18% and88.59%,compared with VGG19 recognition rate,Res Net50,Inception V3 have varying degrees of increase.