多层特征图信息融合的海滩小目标检测

Detection of Beach Small Object Based on Multi-layer Feature Map Information Fusion

提出一种多层特征图信息融合的海滩小目标检测方法,从上下文信息与强化特征图信息融合的角度提升小目标游客的检出率。首先,透过更全面、有效的GAM注意力机制思想结合CSP结构提出GCSAM结构,用于增强检测YOLOv5模型中主干网络跨纬度感受区,聚焦小目标特征学习;其次,在颈部融合端使用BIFPN结构优化YOLOv5网络中PANet结构,补全跨层特征信息之间的传递,使得特征图包含更多的上下文信息;最后,采用幂变换改进YOLOv5网络中CIOU_Loss为Alpha-CIOU_Loss,有效提升预测框的回归精度。实验证明,在满足实时性要求的前提下,相比于原始YOLOv5网络,文中方法在海滩小目标游客检测上查准率提升2.00%,查全率提升5.33%,平均精度均值提升4.36%。文中方法在海滩小目标游客密集、遮挡、目标更小的情况下具有更好的鲁棒性。

Aiming at the problem of insufficient detection accuracy of complex beach small object tourists,a beach small object tourist detection method based on multi-layer feature map information fu-sion was proposed,and the context and feature map information was used to improve the detection rate of small object tourists.Firstly,GCSAM structure was proposed by combining the more compre-hensive and effective GAM attention mechanism idea with CSP structure.The GCSAM structure was used to enhance the cross-latitude receptive field of YOLOv5 backbone.The backbone network was focused on small object feature learning.Secondly,the PANet structure was replaced by BIFPN structure in YOLOv5 network to complete transmission of feature information across layers at the neck network,more context was included in the feature map.Finally,the power transform was used to improve CIOU.Loss to Alpha-CIOU_Loss in YOLOv5 network,and the regression accuracy of prediction frame was effectively improved.Experimental results showed that in comparison with original YOLOv5 network,the precision was improved by 2.00%,the recall was improved by 5.33%,and the mean average precision was improved by 4.36%with real-time requirements.Moreover,this method had better robustness in the case of dense tourists,occlusion,and smaller objects.