Research On Improved SSD Algorithm Based On Feature Enhanced Significantly

As an important research branch in the field of computer vision,object detection has a wide range of applications in many fields.This thesis focuses on the analysis and improvement of the performance of the object detection algorithm based on deep learning,and uses SSD(Single Shot Multi Box Detector)as the benchmark research model to analyze and improve the problems of low performance of model detection and objects including smallscale targets are prone to false detection and omissive detection,which exists in common object detection algorithms represented by SSD.The main research work is as follows:(1)Aiming at the problem of low overall detection accuracy in SSD,it is found that the predictive feature layer does not have the saliency information of different positions,and a object detection model based on cross-dimensional interaction of attention is proposed.For each feature layer,the transformation between different dimensions is completed through a rotation operation,and a spatial attention mechanism is applied to features of different dimensions,and saliency weights are assigned to different positions of the feature layer,thereby enhancing key information while suppressing useless information,then concatenating each attention branch to completing the cross-dimensional interaction of attention.Experiments prove that the object detector integrated attention cross-dimensional interaction can fully focus on the image information that is conducive to the detection task,and improve the overall detection accuracy of the model.(2)Aiming at the problem of misdetection and missed detection of various scales of objects including small objects in SSD,it is found that the feature information of each prediction layer has not been interacted,and the feature information is relatively insufficient,a feature enhanced significantly scheme is designed.Multiple convolution branches including void convolution and attention branches are designed,and multiple branches are concatenated to implement the multi-branch enhancement strategy,so that the prediction layer has a higher level of strong semantic information,and the experiment proves the effectiveness of the scheme.Subsequently,the feature enhanced significantly scheme was integrated into the SSD network,and a Feature Enhanced Significantly Detector(FESD)based on feature enhanced significantly was proposed,the model in this thesis is compared with other algorithm models on the Pascal VOC data set and the traffic signs data set CCTSDB.The experiment found that the model in this thesis respectively achieved 80.7% and 93.8% of the m AP on the VOC2007 and CCTSDB test sets,which not only further improves the model detection performance,but also effectively reduces the false detection and missed detection rate,and proves the effectiveness and generalization performance of the model FESD in this paper.