Author Affiliations
College of Big Data and Information Engineering, Guizhou University, Guiyang , Guizhou 550025, Chinashow less
Fig. 2. Structure of spatial pyramid pooling
Fig. 3. Improved YOLOv3 Model
Fig. 4. Comparison of three situations
Fig. 5. Training loss curve
Fig. 6. Comparison of detection effects of four algorithms. (a) Improved YOLOv3; (b) YOLOv3; (c) YOLOv3-Tiny; (d) YOLOv4
Fig. 7. Overall work flow diagram of system
| Type | Number of filters | Size | Output |
---|
| Convolutional | 32 | 3×3 | 416×416 | | Convolutional | 64 | 3×3 /2 | 208×208 | 1× | Convolutional | 32 | 1×1 | | Convolutional | 64 | 3×3 | | Residual | | | 208×208 | | Convolutional | 128 | 3×3/2 | 104×104 | 2× | Convolutional | 64 | 1×1 | | Convolutional | 128 | 3×3 | | Residual | | | 104×104 | | Convolutional | 256 | 3×3/2 | 52×52 | 8× | Convolutional | 128 | 1×1 | | Convolutional | 256 | 3×3 | | Residual | | | 52×52 | | Convolutional | 512 | 3×3/2 | 26×26 | 8× | Convolutional | 256 | 1×1 | | Convolutional | 512 | 3×3 | | Residual | | | 26×26 | | Convolutional | 1024 | 3×3/2 | 13×13 | 4× | Convolutional | 512 | 1×1 | | Convolutional | 1024 | 3×3 | | Residual | | | 13×13 | | Avgpool | | Global | | | Connected | | 1000 | | | Softmax | | | |
|
Table 1. Network structure of Darknet53
Input | Operator | Exp size | #Out | SE | NL | Step |
---|
4164163 | Conv2d | | 16 | | HS | 2 | 20820816 | bneck,33 | 16 | 16 | | RE | 1 | 20820816 | bneck,33 | 64 | 24 | | RE | 2 | 10410424 | bneck,33 | 72 | 24 | | RE | 1 | 10410424 | bneck,55 | 72 | 40 | 0 | RE | 2 | 525240 | bneck,55 | 120 | 40 | | RE | 1 | 525240 | bneck,55 | 120 | 40 | | RE | 1 | 525240 | bneck,33 | 240 | 80 | | HS | 2 | 262680 | bneck,33 | 200 | 80 | 0 | HS | 1 | 262680 | bneck,33 | 184 | 80 | | HS | 1 | 262680 | bneck,33 | 184 | 80 | | HS | 1 | 262680 | bneck,33 | 480 | 112 | √ | HS | 1 | 2626112 | bneck,33 | 672 | 112 | √ | HS | 1 | 2626112 | bneck,55 | 672 | 160 | √ | HS | 2 | 1313160 | bneck,55 | 960 | 160 | √ | HS | 1 | 1313160 | bneck,55 | 960 | 160 | √ | HS | 1 | 1313160 | Conv2d,11 | | 960 | | HS | 1 | 1313960 | pool,77 | | | | | 1 | 11960 | Conv2d,11 NBN | | 1280 | | HS | 1 | 111280 | Conv2d,11 NBN | | k | | | 1 |
|
Table 2. Network structure of MobileNetv3-Large
[22] Garbage category | Number |
---|
Cigarette case | 460 | Glass bottles | 580 | Milk box | 720 | Plastic bottle | 810 | Screw | 370 | Batteries | 330 | Cigarette end | 420 | Disposable paper cup | 650 | Packing bag | 460 | Pen | 280 | Tissue paper ball | 310 | Toothpick | 650 | Walnut shell | 390 | Banana peel | 410 | Vegetable leaf | 360 | Egg shell | 390 | Orange peel | 420 | Shells of Sunflower seed | 410 | Tea leaves | 420 | Cotton swab | 370 | Drugs | 360 | Lipstick | 530 | Mask | 850 |
|
Table 3. Quantity of various garbage
Network | mAP /% | Model size / MB | Detection speed /(frame·s-1) |
---|
MobileNetv3 | 68.8 | 26.1 | 76 | MobileNetv3+SPP | 70.9 | 27.6 | 74 | YOLOv3 | 67.2 | 246.8 | 43 |
|
Table 4. Performance comparison of different networks
Network | mAP /% |
---|
MobileNetv3+SPP+CIOU | 72.1 | YOLOv3+CIOU | 69.3 | MobileNetv3+SPP | 70.9 | YOLOv3 | 67.2 |
|
Table 5. Comparison before and after the introduction of CIOU function
Garbage category | Improved YOLOv3 | YOLOv3 | YOLOv3-Tiny | YOLOv4 |
---|
Cigarette case | 59.58 | 53.85 | 42.01 | 60.9 | Glass bottles | 69.85 | 61.84 | 54.03 | 73.94 | Milk box | 70.06 | 59.84 | 50.18 | 72.22 | Plastic bottle | 79.98 | 79.91 | 66.86 | 79.97 | Screw | 63.95 | 46.27 | 37.42 | 60.19 | Batteries | 76.21 | 71.21 | 52.57 | 78.7 | Cigarette end | 74.63 | 72.44 | 66.32 | 77.55 | Disposable paper cup | 75.52 | 72.64 | 67.56 | 80.45 | Packing bag | 77.95 | 70.95 | 62.12 | 78.35 | Pen | 74.49 | 73.28 | 60.58 | 76.5 | Tissue paper ball | 73.2 | 70.1 | 55.01 | 74.24 | Toothpick | 76.98 | 74.61 | 64.4 | 77.11 | Walnut shell | 57.6 | 51.57 | 30.53 | 60.09 | Banana peel | 78.85 | 75.75 | 66.88 | 79.85 | Vegetable leaf | 66.14 | 62.68 | 47.49 | 71.8 | Egg shell | 71.95 | 71.52 | 56.49 | 72.8 | Orange peel | 75.35 | 65.25 | 63.28 | 75.87 | Shells of Sunflower seed | 75.51 | 73.88 | 64.5 | 76.61 | Tea leaves | 72.66 | 70.43 | 57.57 | 73.22 | Cotton swab | 65.97 | 64.47 | 63.39 | 68.64 | Drugs | 75.27 | 71.25 | 67.53 | 77.95 | Lipstick | 77.93 | 74.62 | 61.33 | 78.37 | Mask | 68.8 | 68.37 | 53.53 | 69.94 |
|
Table 6. Comparison of AP values of four algorithms
Network | mAP /% | Model size /MB | Detection speed /(frame·s-1) |
---|
Improved YOLOv3 | 72.1 | 27.6 | 74 | YOLOv3 | 67.2 | 246.8 | 43 | YOLOv3-Tiny | 56.6 | 34.9 | 77 | YOLOv4 | 73.7 | 256.6 | 50 | SSD | 65.4 | 97.3 | 48 | MobileNet+YOLOv3 | 66.8 | 25.3 | 75 | MobileNetv3+YOLOv3-Tiny | 55.8 | 11.7 | 82 |
|
Table 7. Comparison of overall detection performance of seven algorithms
Network | Detection speed /(frame·s-1) | mAP /% |
---|
Improved YOLOv3 | 19 | 72.1 | YOLOv3 | 8 | 67.2 | YOLOv3-Tiny | 20 | 56.6 | YOLOv4 | 11 | 73.7 |
|
Table 8. Comparison of detection speed of four algorithms on NVIDIA Xavier NX