Research On Deep Learning Based Fine-Grained Image Classification Method

In computer vision field,the target of traditional image classification task is the objects belonging to different basic categories.Fine-grained image classification has become one of the research hotspots due to the actual application requirements.The target objects of fine-grained image classification often belong to sub-categories which are from one common basic category,and it is a challenging task due to the small difference in inter-class features while intra-class variations of images are large.Besides,complex background features also bring interference to fine-grained image classification task.Based on the research foreground and task difficulty of fine-grained image classification,this paper has constructed and analyzed recognition models for different fine-grained image datasets from two aspects: the loss function improvement during the training stage of convolutional neural network and the refinement of feature extraction process for fine-grained image.The main works of this paper are as follows:(1)Fine-grained plant image classification method based on taxonomic loss.Cross-entropy is the most commonly used loss function in classification tasks of neural networks.Combining cross-entropy loss function and the structure of plant taxonomic hierarchy,this paper presented taxonomic loss for multilevel classification of fine-grained images,which relied on the excellent feature extraction capabilities of convolutional neural networks.The experimental results on Plant CLEF,the plant image datasets in natural background,validated that the taxonomic loss could effectively guide the training process of convolutional neural networks,and it was easy to be implemented.Besides,the comparison results showed that the proposed taxonomic loss outperformed the existing state-of-art methods.(2)Fine-grained rock image classification method based on multiscale feature fusion.In order to integrate the global features and local features of rock image in the training process of neural network,a rock classification method based on multi-scale feature fusion was proposed in this paper,which combined object detection algorithm,multi-scale feature sampling and feature fusion by super-image.The experimental results on Rock-314 showed that the proposed recognition method refined the feature extraction process of rock images,which could fully exploit both multi-scale detail features and global features of rock images.The proposed rock image classification method could effectively capture the finely grained features of images and improve the identification accuracy of fine-grained rock images.