Research On The Application Of Object Detection Technology Based On Deep Learning Algorithm

Object detection,containing image classification and object location,is an important research area in computer vision and the basis of many pattern recognition tasks.With the rapid development of deep learning in recent years,a great breakthrough has been made in object detection based on deep learning,which becomes the main algorithms in the field of obj ect detection.Nowadays object detection technology is widely used in various fields,such as Face recognition,AI medical,self-driving and so on.In this paper,a corresponding improvement method is proposed to optimize small object detection and region proposal network in the two-stage object detection algorithm based on deep learning.The main research contents and innovation points are as follows:1)In this paper,object detection technology is applied to the detection task of aerial remote sensing,and an optimization model based on Faster R-CNN is proposed,which makes use of feature pyramid network to detect and cluster algorithm to help design the sizes of anchors.This model detects independently on different level features and makes use of the prior knowledge provided by clustering to design the sizes of anchors,which greatly improved the detection accuracy of small objects in the task of aerial remote sensing detection.2)Aiming at the problem of imbalance between positive and negative samples in region proposal network training,focal loss is adopted to improve the cross entropy loss.It not only reduces the influence of unbalanced sample number,but also the influence of a large number of easy negatives on the gradient of back propagation in the training process,and strengthens the attention to the hard samples.Experiments show that this method can improve the accuracy of object detection algorithm.3)Aiming at the occlusion problem in object detection,the soft-NMS is used to replace the traditional NMS to improve the way of two-stage detection algorithm to generate region of interest and remove the duplicate box.The suppressed boxes in the original NMS are retained by reducing their confidences,reducing the probability of missing occlusion target.Experimental results show that the detection performance of the optimized algorithm is improved obviously in the serious occlusion categories.