Author Affiliations
1Key Laboratory of Advanced Perception and Intelligent Control of High-End Equipment, Ministry of Education, Anhui Polytechnic University, Wuhu 241000, Anhui , China2Chery New Energy Automobile Co., Ltd., Wuhu 241000, Anhui , Chinashow less
Fig. 1. Network structure of Tiny YOLOv4
Fig. 2. Improved network structure
Fig. 3. Network structures of traditional convolution and depthwise separable convolution. (a) Network structure of traditional convolution; (b) network structure of depthwise separable convolution
Fig. 4. Channel attention structure
Fig. 5. Structure diagram of spatial attention module
Fig. 6. Feature enhancement module
Fig. 7. Loss values of different network structures. (a) Loss values of YOLOv4; (b) loss values of Tiny YOLOv4; (c) loss values of improved network
Fig. 8. Experimental results
Detection algorithm | Number of training parameters |
---|
YOLOv4 | 64040001 | Tiny YOLOv4 | 5939804 | Proposed Tiny YOLOv4 | 4369606 |
|
Table 1. Comparison of training parameters of different network models
Detection algorithm | Model size /MB |
---|
YOLOv4 | 246 | Tiny YOLOv4 | 22.7 | Proposed Tiny YOLOv4 | 15.9 |
|
Table 2. Size comparison of different network models
Dataset | Algorithm | AP /% | FPS /(frame·s-1) |
---|
INRIA | YOLOv4 | 86.2 | 20.1 | Tiny YOLOv4 | 68.67 | 31.6 | Proposed Tiny YOLOv4 | 76.32 | 30.4 | COCO | YOLOv4 | 90.76 | 20.8 | Tiny YOLOv4 | 69.76 | 32.4 | Proposed Tiny YOLOv4 | 78.2 | 30.2 | VOC | YOLOv4 | 90.4 | 19.8 | Tiny YOLOv4 | 71.27 | 31.2 | Proposed Tiny YOLOv4 | 78.8 | 30.6 | Mixed data | YOLOv4 | 92.5 | 19.5 | Tiny YOLOv4 | 73.42 | 32.2 | Proposed Tiny YOLOv4 | 80.52 | 31.4 |
|
Table 3. Test results of different algorithms on different datasets
Algorithm | AP /% | Recall /% | FPS /(frame·s-1) |
---|
Faster R-CNN | 72.37 | 78.4 | 4.7 | SSD | 78.20 | 79.1 | 22.1 | YOLOv3 | 85.34 | 77.6 | 17.3 | Tiny YOLOv3 | 67.80 | 73.4 | 25.6 | YOLOv4 | 92.50 | 81.5 | 19.5 | Tiny YOLOv4 | 73.42 | 75.7 | 32.2 | Proposed Tiny YOLOv4 | 80.52 | 82.3 | 31.4 |
|
Table 4. Comparison of results of detection algorithm
Tiny YOLOv4 baseline | DSC | Attention mechanism | Scale enhancement | AP /% |
---|
√ | | | | 73.42 | √ | √ | | | 74.32 | √ | √ | √ | | 77.27 | √ | √ | √ | √ | 80.52 |
|
Table 5. Ablation experiments on mixed datasets