Research On Object Detection Based On Deep Convolutional Neural Network

Object detection is widely utilized in image retrieval,video surveillance,military reconnaissance and other fields.Its task is to automatically identify classification and location information of the object from complex scenes.In view of the poor portability of traditional modeling methods,deep learning has become the main means of research object detection.In order to improve the accuracy of object detection based on deep learning,this paper researches the object classification and location as the main breakthrough points:(1)The object detection method based on bidirectional feature pyramid is proposed which is aiming at the problem that the FPN(Feature Pyramid Network,FPN)can only integrate high-level information into the low-level information,affecting the classification and location of the object.Firstly,the image is preprocessed and expanded to a uniform size.Secondly,the FPN is used to generate and fuse the multi-scale feature maps,and then the bidirectional pyramid is used to connect the feature maps with a bottom-up reverse side connection method.Then,the connected feature maps are input to the RPN(Regional Proposal Network,RPN)and RoIAlign Pooling(Region of Interesting Align Pooling,RoIAlign Pooling)respectively.The RPN extracts the bounding box and then inputs RoIAlign.Finally,the network is adjusted according to the loss function.The method effectively combines the information of the upper layer and the bottom layer in two directions to successfully solve the problem of inaccurate classification and positioning which is caused by the one-way fusion of the feature pyramid,and improves the accuracy of the object detection.(2)The object detection method based on multi-threshold iterative region extraction is proposed which is aiming at the problem that the RPN prior to the identification of the anchor frame and the background adopts a single threshold to cause the misdetection frame and the over-fitted and the accuracy improvement limited.Firstly,the image is preprocessed.Secondly,the shared convolutional layer is used to extract features,and then the bidirectional feature pyramid is used to bidirectionally fuse the information of the high and low layer feature maps.Then,the feature maps are input into the multi-threshold iterative region proposal network and RoIAlign.The multi-threshold iterative region proposal that the network distinguishes its front and background by setting multiple thresholds for the anchor box,and classifies and returns the foreground.Finally,the proposal box of the multi-threshold iterative region proposal network output is input to RoIAlign,and the extracted the object area and information are regressed and classified.The method effectively distinguishes the front and background of the anchor point frame,and successfully solves the problem that the single threshold causes the positioning to be inaccurate,so that the position of the detection object is more accurate.