Research On Scene Recognition Based On Local Perception

With the rapid development of computer hardware,software and Internet technology,digital images,as a medium of information transmission and storage,have increased by an order of magnitude.How to make the computer automatically understand the image content,and then effectively classify,manage,retrieve and recommend it has become an urgent problem in academia and industry.Scene image recognition has also become a hot research topic in the field of computer vision.Scene is composed of objects in different spatial locations.However,at present,most scene recognition methods are based on the global image information and ignore the importance of local objects.A few scene object-based recognition methods often generate a large number of regional proposals,increasing the computational pressure and recognition time.Therefore,based on local perception,this thesis makes an in-depth exploration of scene recognition tasks.The main works of the thesis are as follows:(1)This thesis proposes a scene image recognition method based on essential scene sub-graph.Through observation,we find that the objects in the same category of scenes have similarities.Therefore,this method firstly analyzes the structure of each kind of scene graph and mines the essential scene sub-graph,then uses the essential scene subgraph and the global scene information to learn a bi-enhanced knowledge space iteratively,and finally utilizes this space to classify scene images.Experimental results show that the algorithm is effective and improves the scene classification accuracy.(2)This thesis designs a scene image recognition model named RMAN based on the recurrent and memorized attention mechanism.Different from the previous methods with strong supervision,RMAN pays attention to different scene areas through the attention localization module,and finally uses the LSTM network to memorize features and classify them.This algorithm realizes end-to-end scene recognition and performs well on several datasets.