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

Fine-grained image recognition is a branch of computer vision that has attracted much attention,and it has extremely important value in academia and industry.In recent years,deep learning has been studied by many scholars and has been frequently applied in more and more fields,and new major breakthroughs have been made.The combination of coarsegrained recognition and ordinary image recognition has achieved unprecedented depth of development.The field of traditional coarse-grained image recognition combined with deep learning has also been greatly developed,reaching unprecedented precision.However,due to fine-grained image recognition,the difference between the different sub-categories of the target species is extremely subtle,and the species of the same sub-category may be affected by various aspects such as illumination,occlusion,posture,and complex background,resulting in a large difference in the appearance of the image.It is also susceptible to interference from information irrelevant to the recognition,so the traditional convolutional neural network alone cannot distinguish fine-grained images well at present,and its recognition accuracy still has great potential to be improved.Therefore,in order to obtain higher fine-grained recognition accuracy,it is necessary to find ways to accurately locate the target object and capture more discriminative detail areas.In order to effectively improve the accuracy of fine-grained image recognition,this paper is based on the deep learning algorithm to study its application effects in the field of fine-grained image recognition.The specific innovations and work are as follows:(1)Many existing methods rely on manual labeling too much,which cost a lot of time and energy.To saving costs,this paper proposes a weak supervised learning method based on multi-scale cross-feature fusion.This model can locate the target object from the picture without the aid of part annotations accurately and discard the interference of background noise on the foreground.Next,the most discriminative salient component regions are captured from the target objects,so as to effectively learn the fine-grained features of different scale parts of the target object.Experiments have proved that the algorithm has obtained 87.4%,90.8%,and 94.2% recognition accuracy on the data sets of CUB-200-2011,FGVC-Aircraft,and Stanford Cars,respectively.Compared with the previous excellent weak supervision methods,the recognition accuracy has been further improved.(2)In order to obtain more recognizable features,this paper proposes a fine-grained image recognition algorithm based on adversarial complementary attention enhancement.The algorithm applies an adversarial erase strategy on the attention module,and learns to locate more discriminable parts regions erasing the response region with the largest activation value in the image.The attention module is used to get the most responsive part and erase it.The network is driven to find another discriminative component.After obtaining the the most discriminative part by attention mechanism,the region of most discriminative part is erased,and the network is driven to find another discriminative part.The similarity loss is used to limit the similarity of the two discriminative part,avoid feature overlap and increase feature diversity.Experiments show that the algorithm in this paper has achieved87.5%,92.8%,and 94.4% recognition accuracy on the CUB-200-2011,FGVC-Aircraft,and Stanford Cars data sets,respectively,which effectively improves the recognition accuracy.