Improved Lightweight Apple Object Detection Method Based on YOLOv5s DOI
Yunfei Jia, Chang Liu

Published: Nov. 17, 2023

Fast and accurate object detection is an important challenging task in the context of automated apple harvesting. However, current models have relatively large network parameters. In this work, improvements been made to algorithm based on YOLOv5s, by replacing original model's backbone with ShuffleNetV2 lightweight structure. To compensate for reduction model parameters, paper introduces ECA mechanism SPPF module after modules improve accuracy. Experimental results demonstrate that improved YOLOv5s reduces parameter count 85.6%, decreases computational load 87.3%, achieves a accuracy [email protected] 97.0% detection, which 0.4% higher than model. The frame rate has also 15 FPS compared model, making it suitable practical real-world environments.

Language: Английский

Research on Metallurgical Saw Blade Surface Defect Detection Algorithm Based on SC-YOLOv5 DOI Open Access
Lili Meng, Xi Cui, Ran Liu

et al.

Processes, Journal Year: 2023, Volume and Issue: 11(9), P. 2564 - 2564

Published: Aug. 27, 2023

Under the background of intelligent manufacturing, in order to solve complex problems manual detection metallurgical saw blade defects enterprises, such as real-time detection, false and model being too large deploy, a study on surface defect algorithm based SC-YOLOv5 is proposed. Firstly, SC network built by integrating coordinate attention (CA) into Shufflenet-V2 network, backbone YOLOv5 replaced improve accuracy. Then, SIOU loss function used prediction layer angle problem between frame real frame. Finally, ensure both accuracy speed, lightweight convolution (GSConv) replace ordinary module. The experimental results show that [email protected] improved 88.5%, parameter 31.1M. Compared with original model, calculation amount reduced 56.36%, map value increased 0.021. In addition, overall performance better than SSD YOLOv3 target models. This method not only ensures high rate but also significantly reduces complexity calculation. It meets needs deploying mobile terminals provides an effective reference direction for applications enterprises.

Language: Английский

Citations

5
Unmanned Aerial Vehicles General Aerial Person-Vehicle Recognition Based on Improved YOLOv8s Algorithm DOI Open Access
Zhijian Liu

Computers, materials & continua/Computers, materials & continua (Print), Journal Year: 2024, Volume and Issue: 78(3), P. 3787 - 3803

Published: Jan. 1, 2024

Considering the variations in imaging sizes of unmanned aerial vehicles (UAV) at different photography heights, as well influence factors such light and weather, which can result missed detection false model, this paper presents a comprehensive model based on improved lightweight You Only Look Once version 8s (YOLOv8s) algorithm used natural infrared scenes (L_YOLO).The proposes special feature pyramid network (SFPN) structure substitutes most neck extraction module with Special deformable convolution (SDCN).Moreover, undergoes pruning to eliminate redundant channels.Finally, non-maximum suppression intersection-union ratio minimum point distance (MPDIOU_NMS) has been integrated boxes, validation conducted using dataset Visdrone2019 dataset.The experimental results demonstrate that when number parameters floating-point operations is reduced by 30% 20%, respectively, there 1.2% increase mean average precision threshold 0.5 (mAP(0.5))and 4.8% mAP(0.5:0.95) dataset.Finally, mAP experienced an 12.4%.The accuracy recall rates have seen respective increases 9.2% 3.6%.

Language: Английский

Citations

0
Improved Lightweight Apple Object Detection Method Based on YOLOv5s DOI
Yunfei Jia, Chang Liu

Published: Nov. 17, 2023

Fast and accurate object detection is an important challenging task in the context of automated apple harvesting. However, current models have relatively large network parameters. In this work, improvements been made to algorithm based on YOLOv5s, by replacing original model's backbone with ShuffleNetV2 lightweight structure. To compensate for reduction model parameters, paper introduces ECA mechanism SPPF module after modules improve accuracy. Experimental results demonstrate that improved YOLOv5s reduces parameter count 85.6%, decreases computational load 87.3%, achieves a accuracy [email protected] 97.0% detection, which 0.4% higher than model. The frame rate has also 15 FPS compared model, making it suitable practical real-world environments.

Language: Английский

Citations

0