Research On Fine-grained Image Classification Algorithm Based On Multi Convolution Neural Network Fusion

Computer vision is an important field of artificial intelligence.In the research of computer vision,image classification has always been a classic research direction.The traditional image classification tasks are generally aimed at the classification of large categories such as "cat" and "dog".However,in the real scene,the objects to be classified often come from the subcategory of a traditional category,such as identifying different kinds of dogs under the species of "dog",This kind of research that focuses on the further sub-category division of the category of the target object in the image from the traditional large category is called fine-grained image classification.Because the fine-grained images have the characteristics of great difference in the same subclass and great similarity in different subclasses,the classification difficulty is much more difficult than the traditional coarse-grained image classification.If combined with the influencing factors such as attitude,angle,illumination and complex background in the image,it is more difficult to identify effectively.In view of the above difficulties,this thesis adopts multi model fusion technology to achieve better recognition effect than single model method.At the same time,a dynamic weighted multi model fusion method for fine-grained image classification is innovatively proposed to solve the problems of insufficient generalization ability of single model method and weight solidification of traditional multi model fusion method.The main research contents are as follows:(1)In this thesis,a dynamic weighted multi model fusion method for fine-grained image classification is proposed.Through experimental verification on two kinds of network models based on transfer learning and attention mechanism,the performance of the model is improved as a whole;At the same time,in the process of model training,an adaptive weight adjustment algorithm is proposed.The algorithm will adaptively adjust its weight value based on the accuracy change of the sub model participating in the fusion in each training process,so as to ensure that the whole model reaches the optimal state.In the experiment,two fine-grained image data sets,agricultural data set under complex background and medical data set under simple background,are used respectively,and the method in this thesis is verified from different evaluation angles.The experimental results show that compared with the traditional single model method,the classification effect of this method is obviously improved;Compared with the multi model fusion method with fixed weight,this method can achieve the optimal state faster while improving the classification effect,saving time and computing cost.In addition,through experiments on two data sets from different fields with simple background and complex background respectively,ideal results are obtained,which proves the generalization ability of this method.(2)In this thesis,sequence and exception attention module,selective kernel attention module and efficient channel attention module are introduced into ResNet model respectively,and three attention-ResNet models of SE-ResNet,SK-ResNet and ECAResNet are established.The three attention ResNet models are experimentally verified on the multi model fusion method proposed in this thesis,Similarly,the agricultural data set under complex background and the medical data set under simple background are used to verify the method in this thesis.The experimental results show that,whether compared with the single attention model or the multi model fusion method with fixed weight,this method can achieve the optimal state faster while improving the classification effect,and the multi model fusion method proposed in this thesis has significant advantages in the fine-grained image classification task of complex background,which proves the superiority and practical significance of the method proposed in this thesis.In addition,this thesis uses the original and unmodified official network model(transfer learning model)and the structurally improved network model(attention model)for experiments,which proves that the multi model fusion method proposed in this thesis also has wide applicability in sub model selection.