Research On Fine-grained Image Classification Method Based On Attention Mechanism

Fine-grained image classification,as a subtask of image classification,has a wide range of application values in product quality detection,automatic biodiversity monitoring,smart retail,and smart traffic monitoring.In recent years,domestic and abroad scholars have discovered that deep learning has powerful feature expression capabilities in image classification,and have applied deep learning to the fine-grained image classification.Therefore,many fine-grained image classification algorithms based on deep learning have emerged and have shown quite good classification performance.The fine-grained image classification task,due to the factors such as different types of targets with high similarity,targets of the same category having different poses or angles,and large background interference,faces big challenge to improve model classification performance.Constructing a reasonable network to obtain the superior feature representations and focusing on the local distinguishing region of the image is the key point to improve the effect of fine-grained image classification.Therefore,this thesis,based on a detailed analysis of the classic fine-grained image classification algorithms,aiming at the problem that the classification algorithms were not comprehensive enough to express the features of fine-grained images,aims to optimize the feature extraction ability of the network.The main work completed in this thesis are as follows:1.Introduce the attention mechanism to improve B-CNN fine-grained image classification algorithm.In view of the fact that the features extracted by the B-CNN algorithm in the feature extraction stage cannot effectively capture the local distinguishing regions of the image,the feature fusion for improving the comprehensiveness of feature representation in the B-CNN algorithm is retained,and the network structure is modified to improve the ability of the network to capture distinctive regions.The attention modules are added into the two network branches before feature fusion to highlight the distinctive local regions in the image.Compared with the mainstream algorithms,the effectiveness of the improved algorithm is verified.In addition,in order to improve classification accuracy,data enhancement is used to expand data sets,Experimental results show that the classification accuracy of the improved algorithm has reached 87.7% and 93.1% on the enhanced CUB-200-2011 and Stanford Cars datasets,respectively.2.Introduce the dual attention mechanism to recognize the workpiece defect.Workpiece defect recognition was regarded as a fine-grained image classification task,and a dual attention network was introduced into the B-CNN network.By the self-attention mechanism,the spatial attention module captures the similarity between adjacent pixels to highlight local features and the channel attention module assists the fusion of semantic information between feature maps of different channels.The experimental results on the diode workpiece data set show that the proposed method can complete the diode defect classification task well,and the classification accuracy reaches 96.4%.