Research On Object Detection Based On Multi-layer Features And A New Generation Mechanism Of Candidates

Artificial intelligence is the hot topic in the present,and the understanding and analysis of the intelligent image is one of the important fields.Image intelligence includes image classification,image segmentation,object detection,etc.Object detection is the prerequisite for the understanding of advanced visual tasks,it is used widely in driverless,face detection,smart video monitoring and augmented reality and so on.In recent years,with the emergence of large image data sets and ability of computer computing,the deep learning technology in the field of computer vision has been a huge success,and has been widely used in the field of object detection.However,due to the complex background of the image and the small size of the object,the final detection accuracy is not satisfactory.For the above reasons,this paper mainly studies the detection accuracy of objects.Convolution neural network in the feature extraction often adopt the deepest feature map,the deeper the feature map,the semantic information is richer,but it will be missing some edge feature of the object,it is not helpful for detect the position of the object,therefore,this paper analysis the characteristics of different feature map of convolution neural network,put forward the characteristics of adjacent residual block at the end of the convolution layer for two,for these new generation of candidate box features are extracted,the characteristics of the rich expression to improve the quality of the candidate box,adopted the shallow convolution features at the same time,more conducive to the detection of small objects.In addition,this paper put forward the way of cascade RPN,will generate candidate box to enter the RPN network,the location of the candidate box refinement,further enhance the quality of the candidate box,at the same time,the target candidate box generated in the process,can produce a large number of negative samples,cause the sample proportion is not balanced,the article adopts the method of the maximum inhibition limit the proportion of positive and negative samples,not only speed up the convergence speed,the detection accuracy is improved.In order to further strengthen the detection object classification effect,this article uses method of global context,the entire image convolution characteristics and combined with the characteristics of the candidate box figure,considering each candidate associations between frame and the whole image,therefore,for the final classification result is improved significantly.In this paper,the experiments are mainly carried out on PASCAL VOC public data set,and the proposed improvement methods are tested respectively.Through the experimental results,we can verify that the above improved methods can effectively improve the final detection accuracy.