| In recent years,deep learning has been widely used to solve practical problems for its excellent performance in the fields of image,speech,and natural language processing.Fine-grained image recognition is a subdivision field under image recognition,which mainly refers to the task of image classification for different sub-categories belonging to the same basic category.And it is more challenging than general image recognition tasks as the inter-class variance is small while the intra-class variance is large among fine-grained categories.This paper mainly studies fine-grained image recognition algorithm and the application of these algorithms in the classification of insect morphology.In this paper the key issues of fine-grained image classification are first analyzed.Then an attention based convolutional neural network is proposed for fine-grained image recognition.This model contains several parts: 1.A object and part level attention model based on the spatial-wise attention of feature map in convolutional neural network.And this model can attend to more fine-grained feature by zooming the discriminative region layer by layer through different CNN;2.Combine the channel-wise attention model and the spatial constrained loss function constructed by the spatial relationship between object and part to keep the discriminative and variety of attended region.The experimental results on fine-grained image classification datasets such as CUB200-2011,Stanford Dogs and Stanford Cars demonstrate the power of our proposed model.In addition,this paper also proposes a hierarchical classification network according to the characteristics of fine-grained categories in biological classification.This model can use the structural information among different categories in the process of fine-grained recognition,which can improve the classification accuracy to a certain extent.Finally,for the problem of detecting finegrained categories from images with complex backgrounds,we propose a two-stage scheme that uses the detection network to distinguish the foreground and background,and then utilize a special fine-grained identification network to identify the objects in the bounding box.The experimental results proved the effectiveness of proposed methods.Finally,we apply the above methods to the fine-grained classification of natural butterfly images and pest categories.The experimental results show the high performance of our proposed method in image-based insect morphology classification,which proves the feasibility and effectiveness of our method for dealing with practical problems. |
