| The research on object detection in image processing is very popular at present.There are a series of problems such as low object detection efficiency and low recognition accuracy in traditional image processing.Accompanied by the emergence of deep learning object detection problem got obvious improvement,but there are still some problems: First,in the process of object detection,image because in the process of shooting under natural light to illuminate,shooting distance,shade is difficult to accurately detect causes such as object detection,for the follow-up work such as semantic segmentation also will impact;In addition,as deep learning requires a large number of convolutional neural network operations,image subsampling,pooling and other operations after passing into the neural network will cause the loss of information features of small object objects and make it difficult to accurately detect small objects.Therefore,in order to solve the problem that the object is occluded and difficult to identify and the small size of the object is difficult to detect in the object detection,this thesis proposes two image object detection methods in the natural scene: object detection and recognition method based on cross-dimensional interactive attention mechanism and small object detection and recognition method based on densely connected residual network.The two methods are mainly introduced as follows:(1)Since there are many object classifications in images taken in natural scenes,there are also some interferences in factors that affect object recognition.Inspired by the attention mechanism,it can automatically ignore the information irrelevant to the object in the image processing process,and proposes an object detection and recognition method based on the cross-dimensional interactive attention mechanism: CDI_AM-Det,used for image object detection in natural scenes.First,use CSPDarknet53 in the Yolov4 method as the backbone feature extraction network to extract features;Then,the spatial pyramid pooling(Spatial Pyramid Pooling,SPP)operation connected to the last feature layer is added with void convolution to become a void space Pyramid pooling operation;Then,introduce a crossdimensional interactive attention module at the prediction network,and branch to perform multi-scale prediction to enhance the use of object information;Finally,eliminate redundant candidates through improved non-maximum suppression Frame,get the final object detection and recognition frame.This method is compared with the Yolov4 method through training,verification,and comparison experiments in the PASCOL VOC data set commonly used in object detection,which proves that this method is feasible.(2)Because the images taken in natural scenes are far away from some objects,it is difficult to detect small objects.Inspired by the dense connection network,the information utilization rate between different layers of the network can be enhanced when the network is trained in object detection.Another research method in this thesis is to solve the problem that it is difficult to identify small size objects in object detection,and to propose a small object detection method based on dense connection residual network: DR-YOLOV4.Firstly,dense network connection was made in the first residual block of CSPDark Net53 trunk feature extraction network,so that the object feature information in the low-level network structure could be fully utilized.Then,the RFB module of increasing receptive field is further introduced,which is placed in front of the prediction network to enhance the prediction function of the prediction network.Finally,the object detection method of small object recognition is obtained.The data set used in this method is KITTI data set with most small object sets,and the effectiveness of this method can be proved by comparing with the original YOLOV4 method.In addition,CDI_AM-DET and DR-YOLOV4,two methods proposed in this thesis,are compared in small object detection,and the performance of this method is better. |
PDF Full Text Request