Fine-Grained Image Recognition Method Research Based On Key Region Detection

Human beings acquire a large amount of image information and external knowledge through vision.In recent years,how to accurately identify the corresponding types of objects of interest is a research hotspot in the field of computer vision.Therefore,it is particularly important to use machine learning or deep learning instead of human beings to realize fine-grained image recognition.Aiming at the problem of fine-grained image recognition,this paper summarizes the current research progress of fine-grained image recognition technology at home and abroad,extracts image features using deep learning technology,and improves the existing fine-grained image recognition methods based on deep learning.The main work of this paper is divided into the following two parts:(1)Fine-grained image recognition method based on attention mechanism and multi-scale imagesDifferent from conventional image classification methods,fine-grained image classification methods require more stringent,which requires accurate recognition of a large category of neutron classes and fine image classification.In the whole image,the target object often does not occupy the whole area of the image.It will not only be affected by too many complex,diverse and ineffective background noise.In the deep neural network,the high-level features of the network lose a lot of local information,which leads to inaccurate recognition results.To solve this problem,attention module is introduced to screen background information and generate feature mask to map the key areas of the target object,and Inception module is used to make different features for the uncertain and multiple key features of the target object.Feature extraction makes the features contain both high-level abstract semantic information and low-level local details.Experiments show that compared with the current mainstream methods,this method based on attention mechanism and multi-scale fine-grained image recognition can achieve better recognition accuracy.(2)Fine-grained image recognition method based on circular feature fusion trainingand saliency detectionIn the research of forecasting based on image pyramid longitudinal feature fusion,we often only focus on the final prediction results,but neglect that the horizontal circular feature fusion and retraining in the training process can obtain better feature representation,so how to improve the feature expression ability in the training process has gradually become a research hotspot.Aiming at this problem,a method of retraining feature fusion based on forward network is proposed.In the forward propagation based on deep convolution neural network,the high-level output feature cycle is input into several low-level structures for feature fusion training,the extraction of feature details is helpful to realize fine feature classification and improve the recognition accuracy of fine-grained images.The experimental results show that,this method can effectively improve the recognition accuracy.