Research On Multi Size Object Detection Method For Single Image

The detection method of single image with multi-dimensional,that is,the detection method for traditional size object and the detection method for small object.Object detection is one of the important research directions in the field of computer vision and digital image processing,and it is the basis of high-level vision tasks such as object tracking and action recognition.However,due to the complexity of the real scene,the existing object detection methods are not compatible with real-time and accuracy,and can not give correct recognition results for small objects.In view of the shortcomings of the existing object detection methods,combined with the idea of deep learning and supervised learning,this paper designs the relevant methods to improve the real-time and accuracy of the object detection methods and the detection effect for small objects.The specific research contents are as follows.Firstly,the basic concepts and related knowledge of object detection are briefly introduced,and the related technologies involved in object detection are summarized.Secondly,in view of the current object detection task is real-time and accuracy can’t compatible problems,put forward a way to use efficient feature reuse thought to realize object detection method,this method first convolution neural network is utilized to extract the characteristics of the input image and generate the initial figure,then introduce effective features reuse modules to solve different layers between figure and detecting object receptive field does not match the problem,finally,using the K-means clustering algorithm optimized regression prior box to realize object detection task.Thirdly,a multi-scale feature fusion object detection method based on adaptive sensing field is proposed to solve the problems of low detection accuracy and inaccurate positioning in the current object detection task.The method based on binary classification forecast layer and effective receptive field design in the prediction,adaptive module o f the reception field is added to the layer to enhance the capacity of the backbone network of small object feature extraction,and then construct two-way characteristic figure weighted fusion structure as a network of classification prediction to fully mix multi-scale features,in the end,the binary classification task and multi classification forecasting task combined to achieve object detection task.Finally,in order to verify the accuracy and effectiveness of the two object detection methods,experiments were carried out on PASCAL VOC,MS COCO and self-made data sets,and compared with the existing object detection methods.