An Improved Deep Learning Object Detection Method Combined With Edge Information

Currently,with the continuous development of computer technology and artificial intelligence,computer vision technology is widely used in various fields.Object detection based on deep learning is an important research field in computer vision,and has significant application value in areas such as face recognition,intelligent transportation,and autonomous driving.In deep learning algorithms,the problem of sample imbalance is a hot topic,and the setting rules and sampling methods for negative samples,are key to solving the problem of sample imbalance and largely determine the effectiveness of the detector.Meanwhile,edge detection,as a fundamental task,is closely related to the division of the object to be recognized and the background,but is often overlooked in the object detection process.To address this issue,this paper proposes an edge-guided improved object detection method combining traditional edge detection and deep learning algorithm,aiming to solve the problem of insufficient hard negative samples in sample imbalance and improve the accuracy and robustness of object detection.Firstly,to overcome the limitations of traditional edge detection algorithms,this paper proposes a multi-scale edge detection method.Based on a topological analysis method of binary image edge tracking,we uses channel decomposition,multi-threshold binarization,multikernel Gaussian filtering,and multi-scale clustering methods to generate more comprehensive,accurate,and robust results.By mapping the results to rectangular boxes,it achieves effective application in the scene of deep learning object detection.Secondly,to solve the problem of insufficient hard negative samples in deep learning algorithms such as Faster R-CNN,we propose an edge-guided module that improves the region proposal network of the Faster R-CNN.By using our multi-scale edge detection method,we improve the sampler and allocator of the region proposal network,propose an edge detectionguided balanced sampling method,which increases the proportion of hard negative samples without increasing the total number of negative samples,thus improving the performance of the RPN and increasing the recall of the proposals on the Microsoft Common Objects in Context dataset by 1.7 percentage points relative to the original.Finally,in order to better apply the improved region proposal network to Faster R-CNN and further improve the detection performance of the network,we use more flexible deformable convolution to improve the convolution layer of the backbone.This makes the network focus more on feature edges during feature extraction,enhances the network’s adaptability and representation ability,and thus learns better from hard negative samples.In addition,due to the increase in hard negative samples in the early stages of training,region proposal network provides inaccurate proposals,which leads to poor convergence in the early stages of the network.To address this issue,this paper improves the network by introducing a Complete Intersection over Union loss that emphasizes more on the edges and shape differences of the object.The improved Faster R-CNN achieves a relative improvement of 5.4 percentage points in mean average precision on the Microsoft Common Objects in Context dataset compared to before,and has higher accuracy compared to other object detection methods.