Research On Fine-grained Image Recognition Method Based On Deep Learning

Fine-grained image classification refers to the finer division of subcategories within a large category,such as a whether a dog is a Husky or a Samoyed.The challenge of fine-grained image recognition lies in the small difference between different categories and sometimes large difference between the same category.Due to the inherent characteristics of fine-grained images,such as small discriminant region and insufficient differentiation of image features,the previous recognition methods have some problems,such as single and incomplete information extraction with discrimination and insufficient feature enrichment,so there is still some room for improvement in this task.Aiming at these bottlenecks,this paper designs several different finegrained image recognition methods based on deep learning,which can extract as much and accurate information as possible from fine-grained images under weakly supervised scenes,so as to improve the recognition accuracy.The work completed in this paper is as demonstrated:1.Efficient channel attention multi-branch network method for Fine-grained image recognition.Aiming at the problem that the existing multi-branch structure network mostly only used single attention suggestion sub-network to capture image details,which results in a relatively single information of image components and affects recognition performance,the method in this chapter is proposed.Firstly,based on the recurrent attention neural network,a multi-branch network based on efficient attention module is designed to capture the interaction information between channels and the location information of the target more accurately.Secondly,depthwise over-parameterized convolution is used to replace common convolution so that the number of parameters that can be learned by the network increases.In addition,the improved attention module is used to cut out several key parts of the image so that the information of the image parts captured by the network is more diversified.Finally,experiments are designed and the final results obtained in multiple fine-grained datasets show that the proposed method can effectively improve fine-grained recognition accuracy.2.Enhanced feature improving method for fine-grained image recognition.Most of the existing methods,when extracting image features,do not extract enough features other than the most significant local features,and deal with local features alone,ignoring the relationship between features.Therefore,based on the feature boosting suppression module,this paper proposes an enhanced feature improving method by adding pyramid residual convolution,softpool method and feature focalization module.In this method,pyramid residual convolution firstly uses convolution kernels of different scales to capture features of different levels in the scene,while softpool method reasonably allocates the weight of information in the pooling process.The combination of the two methods enhances the non-most significant feature extraction ability of the network.Secondly,the feature focalization module uses more features mined from the above process to focus on obtaining similar information in multiple local features as discriminative features,which further improves the recognition effect.Experiments designed on multiple datasets demonstrate the validity and generalization of the proposed method.3.Fine-grained recognition method based on biological phylogenetic tree.Inspired by some fine-grained recognition methods that introduced concepts and tools from other disciplines to improve the recognition effect,firstly,construct a phylogenetic tree for the dataset according to the prior knowledge of biology,and secondly,optimize it according to the characteristics of the dataset to construct a taxonomy tree,so as to construct an auxiliary classification system based on prior knowledge of biological taxonomy tree.Pre-classification reduces the misjudgment rate of different species at higher levels.Then,the deep learning network based on Transformer is introduced to improve the recognition effect.The recognition rate of this experimental method on two different biological image datasets is at an excellent level,indicating the feasibility and efficiency of this method.