Research On Fine-grained Image Recognition Technology Based On Weak Supervised Learning

At present,the image recognition technology based on deep learning has greatly improved the accuracy compared with the traditional image recognition technology.However,general image recognition technology usually only recognizes images with large differences,and the recognition of similar objects is insufficient.Therefore,with the focus on “finding multiple detail parts describing the features of objects”,this thesis proposes a new technique of fine-grained image recognition from object location and attention mechanism.The main research contents are as follows:Firstly,the weakly supervised positioning algorithm is used to find the detailed features.This thesis proposes fine-grained image recognition technology based on weak supervised object localization.Since the most significant area tends to focus too much on the details and lacks the perception of the overall structure of the object,and the information of the sub-significant area is the complement of the most significant area,the combination of the two information can more fully express the overall information of the object.In this thesis,the CAM method is used to locate the most significant area and the sub-significant area,and we integrate the two areas to complete the location task.Experiments show that this method has achieved good results in ILSVRC 2012 Validation.Finally,this thesis uses the detailed features to complete fine-grained image recognition,and has achieved good results on the CUB-200-2011 dataset.Secondly,this thesis uses attention mechanism and proposes an end-to-end network based on multi-significant regions.Firstly,according to the two attention mechanisms of channel weighting and space weighting,two salient modules are highlighted and suppressed.Then we cascade multiple saliency modules to implement a progressive network structure,let the neural network explore a plurality of different saliency regions.Finally,we complete image classification by combining multiple saliency features.The experimental results show that the fine-grained image recognition model based on multisignificant regions can find multiple details and achieve high recognition accuracy and recognition efficiency.