Small Object Detection Based On Auxiliary Descent Handover Ratio Loss Functio

Small object detection is a challenging and popular topic in computer vision,with deep learning achieving breakthroughs and successful applications in fields like national defense,intelligent transportation,and industrial automation.However,factors like small size,deformation,complex background,and blurring lead to feature extraction problems and weak generalization in small object detection.In addition,IoU loss function is not applicable to small objects,causing convergence problems.To address these issues,this paper proposes a loss function suitable for small object detection and a small object detection network based on this loss function.The main research content of this paper is as follows:(1)To address the small object convergence problem,this paper proposes an auxiliary descending intersection over union(Adioc)loss function.In convergence process,the predicted box needs to converge towards the ground truth along the centerline,and the loss should steadily decrease.Based on the analysis,this paper first designs an intersection over convex(IoC)function.The IoC loss avoids the zero gradient problem.The Adioc loss function is established by adding a geometric enhancement factor and is theoretically to have better convergence.When applied to small object detection,results show that the Adioc loss function can improve the model’s detection accuracy and training speed.(2)To address feature extraction problem,this paper constructs a TCSP-Yolo small object detection network based on the Adioc loss and Yolo network.Due to the characteristics of small objects,Yolo has difficulty extracting small object features.This paper designs a fine-grained feature extraction module C3 TB,which can solve the problem.To address object aggregation,this paper integrates Transformer module into the bottleneck to extract the global information and context.Also,the CBAM module is injected into the TCSP-Yolo network to increase the model’s sensitivity to small objects.Finally,a detection head is added to improve the network’s detection accuracy.Experimental results show that the Visdrone2019,MS COCO,and Wider Face dataset all demonstrate the effectiveness of the TCSP-Yolo network.