Laser & Optoelectronics Progress, Volume. 60, Issue 10, 1010010(2023)
Detection Algorithm of Recyclable Garbage Based on Improved YOLOv5s
Figures & Tables(12)
Fig. 4. Structure diagram. (a) Schematic of Neck fusion; (b) schematic of improved YOLOv5s overall network
Fig. 6. Original image and image obtained by modules.(a) Original image; (b) image obtained by CBS module; (c) image obtained by S-b module
Fig. 7. Comparison of indicators between the improved model and YOLOv5s in training process
Table 1. Configuration of backbone network structure and parameters
View tableTable 1. Configuration of backbone network structure and parameters
Serial number n Params Module Configuration Output size 0 1 3520 Focus ［3，32，3］ 32×320×320 1 1 3968 S-b ［32，64，2］ 64×160×160 2 1 2528 S-a ［64，64，1］ 64×160×160 3 1 14080 S-b ［64，128，2］ 128×80×80 4 3 27456 S-a ［128，128，1］ 128×80×80 5 1 52736 S-b ［128，256，2］ 256×40×40 6 3 104064 S-a ［256，256，1］ 256×40×40 7 1 203776 S-b ［256，512，2］ 512×20×20 8 1 656896 SPP ［512，512，5，9，13］ 512×20×20 9 1 134912 S-a ［512，512，1］ 512×20×20
Table 2. Ablation experiment data of general category
View tableTable 2. Ablation experiment data of general category
Backbone Neck P /% R /% mAP_0.5 /% mAP_0.5∶0.95 /% Parameters /106 Memory /MB 89.37 85.18 92.09 68.88 7.09 13.7 √ 88.85 89.43 93.36 68.10 4.08 8.03 √ 88.41 89.15 93.03 71.40 5.69 11.05 √ √ 90.12 89.95 94.01 71.30 2.68 5.34
Table 3. Experimental data for ablation of individual class
View tableTable 3. Experimental data for ablation of individual class
Class P /% R /% AP /% YOLOv5s Proposed model YOLOv5s Proposed model YOLOv5s Proposed model Edible oil barrels 93 94.1 91.5 97.9 97.2 97.4 Pop cans 84.4 90.1 76.3 85.9 89 91.8 Cartons 90.8 91.4 91.7 93.5 96.3 97.3 Metal cans 81.8 80.9 78.9 85.3 84 88.7 Beverage bottles 77.3 81.7 74.4 76.2 82.9 86.1 Plugs and wires 82.6 86.5 67.6 78.2 80.6 88.2 Book and paper 94.7 95 94.6 97.5 98.1 98.2 Pans 95.9 93.3 87.2 91.7 96.4 97.7 Scissors 97.2 90.4 89.6 93.3 97.1 95.2 Bowls 96 97.8 100 100 99.3 99.5
Table 4. Comparison experiments with common models
View tableTable 4. Comparison experiments with common models
Model P /% R /% mAP_0.5 /% Parameter /106 Memory /MB Faster-RCNN（ResNet50） 68.50 94.15 92.32 28.3 108.48 YOLOv3 88.47 74.25 85.83 61.9 235.71 YOLOv4 87.83 85.60 88.50 63.9 244.48 YOLOv5s 89.40 85.20 92.10 7.09 13.70 YOLOx-s 92.13 92.53 94.40 8.94 34.29 Proposed model 90.12 89.95 94.01 2.68 5.34
Table 5. Comparison of processing time at Jeson Nano
View tableTable 5. Comparison of processing time at Jeson Nano
Image size Model Preprocessing time /ms Inference time /ms NMS time /ms FP time /ms 640×640 Yolov5s 1.6 182.3 10.1 194 Proposed model 1.5 160.2 9.9 171.6 512×512 Yolov5s 1.1 123.3 9.2 133.6 Proposed model 1.1 108.2 8.8 118.1 416×416 Yolov5s 0.9 82.9 8.0 91.8 Proposed model 0.9 73.2 7.4 81.5 320×320 Yolov5s 0.8 56.0 6.5 63.3 Proposed model 0.7 48.5 5.6 54.8
Tools
Get Citation
Copy Citation Text
Anneng Luo, Haibin Wan, Zhiwei Si, Tuanfa Qin. Detection Algorithm of Recyclable Garbage Based on Improved YOLOv5s[J]. Laser & Optoelectronics Progress, 2023, 60(10): 1010010
Paper Information
Category: Image Processing
Received: Jan. 21, 2022
Accepted: Mar. 1, 2022
Published Online: May. 10, 2023
The Author Email: Haibin Wan (hbwan@gxu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20