Research On Fine-Grained Object Classification Technology Combining Attention Mechanisms

With the development of deep learning,fine-grained object classification has gradually become a research hotspot.Unlike generic image classification tasks,fine-grained object classification aims to distinguish sub-categories,such as different breeds of dogs.Objects from different sub-categories may have subtle visual differences that are even difficult for humans to recognize,while objects within the same sub-category may have large appearance variations due to complex backgrounds,occlusions and changes in scale and viewpoint.Many existing methods rely on the manual definition and annotation of objects and discriminative parts,which are subjective and difficult to generalize.In view of the characteristics of the fine-grained object classification problem and the deficiencies of the existing methods,this paper presents a method for fine-grained object classification combining attention mechanisms.Firstly,the object to be classified in the image is segmented by a convolutional neural network-based model.Then the bounding box is obtained according to the mask and the image is cropped.Finally,the image patch is classified by a model which combines convolutional neural network,recurrent neural network and attention mechanisms.The effectiveness of the proposed method is demonstrated through experiments on three datasets.