Research On Object Detection Algorithm Based On Deep Learning

With the rapid development of artificial intelligence technology,research in the fields of computer vision and natural language processing has ushered in a period of rapid development.As one of the most challenging problems in the field of computer vision,object detection has received a lot of attention and has also made great progress.There are many current object detection algorithms,such as Faster R-CNN,YOLO-v3,SSD,etc.Among them,the information of the different features of the SSD is independent of the algorithm,and the shallow features lack rich semantic information,which leads to the poor detection ability of the SSD for small objects and missed objects.In response to these problems,this thesis is mainly based on the SSD object detection algorithm to carry out research.The main work is as follows:(1)Aiming at the problem of poor detection results of small objects and missed objects by SSD object detection algorithm,an improved SSD object detection algorithm based on feature fusion,DMSFFD,is proposed.The first feature fusion of DMSFFD is to up-sample the shallow features and then fuse them with the high-level features to enhance the semantic information of the features.After the first feature fusion,the DMSFFD algorithm does not directly perform detection,but performs in-depth fusion of the first fusion result,making full use of the complementarity between feature information and enhancing the ability of feature information expression.The experimental results show that the DMSFFD algorithm can better improve the detection accuracy.Compared with Faster R-CNN,SSD and DSSD,the mAP on the VOC 2007 data set has increased by 15.3%,8%,and 7.4%,respectively;compared with Faster R-Compared with CNN,SSD and DSSD,the mAP on the VOC2012 data set has increased by 17%,9%,and 7.5%respectively.(2)In the above-mentioned DMSFFD method,the feature extraction network adopts the VGG model.ResNet has higher classification accuracy and faster convergence speed.In addition,the introduction of feature fusion in the ResNet model can complete the representation of multi-scale features.SENet can automatically obtain the importance of each feature channel through learning,and then extract useful features according to this importance and suppress features that are not very useful for the current task.We propose an improved SSD object detection algorithm SESSD based on the attention mechanism.Because the amount of parameters of the attention mechanism is minimal,the detection performance can be improved without adding too much calculation.We adopt the ResNet model of the standard FPN structure as the feature extraction network of SESSD.The experimental results show that the SESSD model has a certain degree of improvement in the detection ability of small objects and missed objects on the PASCAL VOC test set.Compared with SSD,DSSD and DMSFFD,the mAP on the VOC2007 data set increased by 10.3%,9.4%,and 2.3%respectively;compared with SSD,DSSD and DMSFFD,the mAP on the VOC2012 data set increased by 11%,9.5%,1.7%.On the KITTI dataset,compared with SSD,Faster R-CNN and YOLOv3,mAP has been improved 1.3%,3.9%,0.6%respectively,and the accuracy of each category has been improved to different degrees.