The Research On Representation Method Of Finegrained Feature Of Multi-scale Distinct Image Based On Adaptive

Fine-grained image recognition is a kind of algorithm task that studies how to distinguish different sub-categories from the same large category.The technical challenge with this algorithm is that species in different categories differ only in the most subtle areas(for example,birds' eyes,birds' mouths,birds' or claws).The species in the same category show great differences due to light,posture or background.Therefore,how to accurately identify the distinguishing regions and extract the rich and effective fine-grained features has become the focus of the fine-grained image recognition task.The problems existing in the current fine-grained image recognition methods are as follows: First,the most discriminating parts are blocked or hard to be presented in reality,which leads to the difficulty in accurate classification;Second,the problem of target scale change.Due to the different proportion of the objects in the sample,the information obtained from the target is of different scales with only the same "perceptive field",so it is difficult to accurately classify.Based on the above problems,this paper proposes an adaptive multi-scale discriminative fine-grained feature representation method,which works as follows:(1)A discriminative fine-grained feature representation method based on dual attention is proposed(DFM).This method activates the region by using the attentional mechanism of the channel and spatial information to generate dual attention,and encourages the network to learn multiple differentiated regions in the sample by using the hiding and highlighting modules.(2)A multi-scale feature fusion method based on adaptive adjustment of receptive field was proposed9(RFAM).In this method,by effectively combining the convolution of expansion with different expansion rates,the multi-scale global context information can be obtained by using the multi-scale "perceptive field" to see the target,and the effective context feature information can be obtained by the adaptive allocation of the weights of different "perceptive field" features by the attention module.Combining the design of the above two methods,this paper conducts experiments in two algorithm frameworks respectively,in which the DFL-CNN + DFM + RFAM model based on VGG16 exceeds the baseline 1.2 on the CUB-200-2011,Stanford Cars and FGVC Aircraft data sets respectively %,1.2% and 1.4%,the NTS-Net + DFM + RFAM model based on Res Net50 exceeded the baseline by 1.0%,1.3% and 1.7% on the CUB-200-2011,Stanford Cars and FGVC Aircraft datasets,respectively.Experimental results show that embedding the method modules proposed in this paper on different basic networks,algorithm frameworks and data sets can improve performance.