Comparative Study of the Effectiveness of Multiple Improved Models of the YOLOv8 Algorithm

GONG Yuxiang; GAO Lin; ZHANG Hao; FU Desu

doi:10.13501/j.cnki.42-1908/n.2024.12.010

Abstract：To address the issue of the numerous and various improvement methods for the you only look once version 8(YOLOv8)algorithm and the lack of object detection performance comparison on a common dataset,the publicly available visual object classes(VOC)(2007+2012)dataset was used as the benchmark.The YOLOv8 nano(YOLOv8n)algorithm,with smaller parameters,was selected as the baseline model.Improvements were compared across five aspects:attention mechanisms,backbone networks,neck networksdetection heads,and loss functions.The best individual improvement modules were identified based on mean average precision(mAP)and detection speed.Further combination experiments yielded two optimal combined models with the highest mAP and fastest detection speeds.The results showed that,compared to the benchmark model,the mAP values of the best backbone networksneck networks,and detection head modules increased by 2.50％(Repvit),1.75％(CGDown),and 1.75％(DyHead),respectively;the detection speed increased by 12.85％(RGCSPELAN),2.60％(WaveletPool)and 20.22％(LSCD)respectively;the best attention mechanism module improved the mAP value by 1.88％(CAFM);the loss function did not improve the mAP value and the detection speed.Compared to the benchmark model,the combined model with the highest mAP improved the mAP value by 3.13％(YOLOv8n+CAFM+CGDown+DyHead),the combined model with the fastest detection speeds improved 31.11％(YOLOv8n+RGCSPELAN+LSCD).The former combined model is suitable for high-precision object detection scenarios,the latter combined model is suitable for deployment in edge computing devices with high real-time requirements.This study provides a reference for the improvement of the YOLOv8 algorithm.