Research On Weak Supervision Fine-grained Image Recognition Technology Based On Deep Learning

As a result of the progress of artificial intelligence theory and the improvement of computer hardware level,computer vision technology has developed rapidly in recent years,and the practical products related to computer vision technology have gradually entered our lives.Image recognition technology is an important research field of computer vision technology,which can be divided into coarse-grained image recognition technology and finegrained image recognition technology according to the recognition scene.In the past few years,most of the research has focused on the coarse-grained image recognition technology,but driven by the practical application scenarios,more and more attention has been paid to the study of fine-grained image recognition technology.Compared with the coarse-grained image recognition technology,the fine-grained image recognition technology is to identify subclass targets under the same category,so the task scene corresponding to this technology is more difficult.Based on the deep learning method,this thesis research the fine-grained image recognition algorithm model which uses weak supervision information to complete the supervised learning,and uses the relevant datasets to verify the performance of the proposed method.The main work of this thesis can be divided into the following three parts:(1)Aiming at the characteristics of fine-grained images,IBN-LANet is designed in this thesis by improving the ability of the algorithm model to extract features of different levels and reducing the interference of image style on the model.IBN-LANet is based on Deep Layer Aggregation network,and the method of instance-batch normalization is introduced in the IBNLANet's normalization layer.Through this design,the network improves the ability to extract detailed features,and has better robustness for different image styles in the input image data,thus improving its ability to recognize fine-grained images.Experimental results on relevant fine-grained image dataset show that the network model achieves good accuracy in fine-grained image recognition.(2)Based on the purpose of improving the feature expression ability of the network and retaining more detailed feature information extracted from the convolution layer of the network,this thesis makes further exploration on the basis of the work content of the first part.Aiming at the requirements of fine-grained image recognition tasks,this thesis analyzes the shortcomings of losing too many details when the data in the convolutional neural network transits from feature map to feature vector,and then on the basis of the first part of this work,the global pooling layer of the network is transformed,and FG-Net and FG-LANet are designed.The two networks are characterized by the addition of the design of the global covariance pooling layer.By using the second-order statistical covariance to replace the first-order statistical mean as the pooling output,the feature information contained in the feature vectors after pooling is greatly enriched,and the expression ability of the network to detail features is improved.In this thesis,the two networks were tested in relevant fine-grained image datasets,and the results show that FG-Net and FG-LANet have reached the excellent recognition level in fine-grained image recognition tasks.(3)In order to further explore the algorithm model designed in this thesis,in addition to the fine-grained image dataset,this thesis also uses ImageNet 2012,a large image dataset,to conduct experiments on the three designed networks.The results show that the model structure design proposed in this thesis also achieves excellent recognition capacity in the image recognition task of 1000 categories.It shows that the designed network structure has good generalization performance and can be further explored in more image recognition tasks.