Object Detection And Segmentation Technology Of Complex Background Based On R-CNN

The main task of object detection and segmentation technology is to identify and locate multiple target objects in images or videos.It has been widely used in many fields such as unmanned driving and military analysis,and has is one of the core tasks in the field of computer vision.At present,the research on object detection and segmentation technology has made breakthrough progress.However,object detection and segmentation under complex backgrounds still face many challenges: 1)A small target is a target with fewer pixels in the image,and often has problems such as inconspicuous features,low resolution and susceptibility to background interference,making the detection of small targets unsatisfactory;2)Multiple targets in complex backgrounds often have similar targets,occlusions,target deformations,etc.,increasing the probability of multiple targets failing to detect or miss detect.Therefore,an in-depth analysis of object detection and segmentation algorithms under complex backgrounds is very difficult and meaningful research.Given the above problems,this article started with practical applications and combined R-CNN algorithm in Deep Learning to study target detection and segmentation techniques under complex backgrounds.The main research work and innovations are as follows:Firstly,this paper summarizes the classical target detection and segmentation algorithms in detail,including the traditional machine learning and deep learning-based target detection and segmentation algorithms,as well as the related technologies,which provide the theoretical basis for the research content of this paper.Then,in order to solve the problem that the existing target detection algorithms have low detection accuracy for small targets,this paper studied the Faster R-CNN algorithm,and on this basis,a detection algorithm suitable for small targets is proposed.This algorithm firstly aimed at the problem that it is difficult to extract the features of small and medium objects in the image and designed a multi-layer feature fusion method,which fuses the low-level detail information and high-level semantic information of the target by adding the feature graphs,so as to extract the features of small objects more accurately.Second,a new candidate region proposal network structure—multi-scale RPN is designed to generate small target initial candidate regions with more accurate locations.To solve the problems of imbalance of positive and negative samples and difficult regression of small target bounding boxes during training,a positive sample perceptual loss function focusing on positive samples in the training data of mining data and a boundary regression loss function based on GIo U is designed.Finally,experimental verification is performed on the self-constructed catenary parts detection data set and public data set VOC2007.The experimental results show that the improved algorithm in this paper has higher accuracy for small targets in complex backgrounds,and provides a feasible direction for small target detection.Finally,this paper proposed a multi-object segmentation algorithm based on improved Mask R-CNN for the problems of deformation,similarity and mutual occlusion between multiple targets in a complex background.The algorithm first added a channel attention module to the backbone feature extraction network,which can assign different weights to the features of different targets in the image by self-learning,so as to obtain more effective target features.At the same time,the hybrid dilated convolution was used to replace the partial pooling operation in the feature extraction network.The hybrid dilated convolution can increase the receptive field without changing the resolution of the feature image,thus reducing the information loss caused by the pooling and deconvolution process,and ensuring the integrity of the global feature.Then,a more accurate region of interest pooling operation---Precise Ro I Pooling,was used to solve the problem that the number of interpolation points in the Ro I Align module needs to be set manually,and the size of the feature map cannot be adjusted adaptively.Aiming at the case where Mask R-CNN is not ideal for target edge segmentation,an edge detection branch was designed to be added to the end of the network as an auxiliary network,and the edge detection result was used as a new loss function of the network.Finally,experiments performed on the public dataset Cityscapes.The experiment results show that the algorithm used in this paper can effectively detect and segment multiple targets in complex backgrounds.In summary,this paper focused on the problem of small target detection and multi-target segmentation in complex scenes and proposed corresponding improvements to the problem.Through a series of comparative experiments,it is shown that the method in this paper improves the accuracy and efficiency of small target detection and multi-target segmentation in complex backgrounds,and has some reference application value.