Research And Implementation Of Fine-Grained Image Classification Algorithm Via Weakly Supervised Saliency

The fine-grained image classification task is a sub-area of rapid development in the field of computer vision research,which aims to distinguish small differences between various subordinate categories,such as distinguishing different vehicle models and different kinds of birds.Because fine-grained images are mainly characterized by small differences between sub-categories and large intra-class differences,the current research relies heavily on the annotation of key points or local annotation information,which increases the dependence on manual data annotation,limits the feasibility of practical applications.Therefore,this paper focuses on weakly supervised methods,without additional manual labeling information.The main research is as follows:(1)In order to verify the significance of the saliency map in the depth neural network classification accuracy,a two-branch network fusion structure is proposed.One branch extracts the original image feature and the other branch uses the corresponding saliency map,these images are pre-computed and given as inputs.After the fusion layer is merged,Inception_V3 is used as the basic feature extraction network model,and the CUB-200-2011 data set is verified,the accuracy of fine-grained image classification is improved by about 5%.(2)In the absence of additional object location information or key point information,using the weakly supervised method,the Grad-CAM algorithm is used to obtain the saliency map of the convolution layer of the last layer of the CNN network structure.Aiming at the saliency map,this paper proposes a new local key area localization method,which can accurately locate and extract the most differentiated local areas and reduce the dependence of manual annotation information.(3)In order to make better use of global features and local feature information,this paper proposes a bilinear depth neural network based on Inception_V3.The upper layer network is responsible for extracting the features of the local regions of the image.The lower layer network is responsible for extracting the global features of the image.The global and local features are jointly trained,and the loss function of the hierarchy is constructed to alternately train the model.The publicized fine-grained data set is verified,the results show that the proposed model is close to the latest relevant research results in classification performance,and it can have good classification accuracy in different data sets and has certain robustness.In the actual vehicle inspection project,the classification results for the vehicle type,vehicle brand and vehicle model have obtained more than 90% accuracy,and have strong practicability.Based on the weak supervision method and reducing the dependency of manual labeling,this paper proposes a new fine-grained image classification model.Although it has achieved good classification results in the public data set,it needs to further realize end-to-end training,and the thousand target categories need to further improve the accuracy of the classification.