Research On Object Detection Algorithm Under Low Light Based On YOLOX

Object detection aims to achieve the goal of sensing the scene by classifying and locating the objects in the image,so that machines can identify objects like humans and other similar organisms.In the field of computer vision,object detection is at the core,with important and widespread applications,and has been under intense research,even becoming the foundation of other research fields.However,in low light scenes,object detection is difficult and research is relatively limited,and the detection results are greatly affected by the detection background.Although various night assistance systems can be used to improve the shooting effect and indirectly improve the detection results,this widespread use of hardware facilities requires significant investment.In response to the above issues,this article has studied various classic object detection algorithm models based on deep learning according to the timeline,and selected the YOLOX algorithm of Megvii Technology Limited as the research basis and made certain improvements to it.Starting from the algorithm itself,it directly improves the object detection effect in low light fields.The main research contents of this paper are as follows:(1)Improvement of attention mechanism.By adding multiple attention mechanisms to the output of the model’s backbone network,the model significantly places weight on important features before entering the Neck section,finds important features on the original data through correlation,and makes these important features prominent in the network,enabling the network model to better complete classification and positioning.(2)Improvement of loss function.Adding a modulation factor to the binary cross entropy loss function in the confidence prediction loss function section balances the proportion of positive and negative sample losses to the total loss,optimizing the network model in the desired direction and better classifying positive samples.(3)Improvement of model structure.An adaptive spatial feature fusion(ASFF)module is added to the tail of the PAFPN.The network filters out conflicts between information through the ASFF strategy,thereby preventing inconsistent feature ratios during gradient backpropagation due to multi-scale changes,while reducing the overhead of reasoning and computing,and improving the accuracy of the network model.In this paper,the structure and confidence loss function of the YOLOX algorithm have been improved.Through comparative analysis of experimental results,the detection accuracy of the YOLOX algorithm in low light scenarios has been significantly improved,which proves the feasibility of this scheme and has certain value for practical applications of object detection.