Author Affiliations
1Flat Panel Display National and Local Joint Engineering Laboratory, National University Science Park Sunshine Technology Building, Fuzhou University, Fuzhou, Fujian 350116, China2College of Physics and Information Engineering, Fuzhou University, Fuzhou, Fujian 350116, China3Jinjiang Bosen Electronic Technology Co., Ltd., Quanzhou, Fujian 362200, Chinashow less
Fig. 1. YOLOv3 network structure
Fig. 2. Structural unit of YOLOv3. (a) Convolutional unit; (b) ResNet unit; (c) Convolutional Set unit
Fig. 3. Improved backbone network
Fig. 4. Improved dense network module
Fig. 5. Improved spatial pyramid pooling network module
Fig. 6. DS-YOLO network structure. (a) Network structure; (b) DC-SPP unit; (c) Conv-Residual unit
Fig. 7. YOLOv3tiny test results
Fig. 8. DS-YOLO test results
Fig. 9. YOLOv3 test results
Method | Weight /MB | BFLOPS | mAP /% | Speed /(frame/s) |
---|
Base-YOLO | 17.2 | 7.8 | 65.9 | 150 | YOLOv3 | 236 | 65.86 | 78.3 | 30 |
|
Table 1. Comparison of Base-YOLO and YOLOv3
Method | Weight /MB | BFLOPS | mAP /% | Speed /(frame/s) |
---|
Base-YOLO | 17.2 | 7.8 | 65.9 | 150 | D-YOLO | 24.6 | 8.54 | 67.7 | 143 |
|
Table 2. Comparison of D-YOLO and Base-YOLO
Method | Weight /MB | BFLOPS | mAP /% | Speed /(frame/s) |
---|
Base-YOLO | 17.2 | 7.8 | 65.9 | 150 | S-YOLO | 17.5 | 7.92 | 67.1 | 147 |
|
Table 3. Comparison of S-YOLO and Base-YOLO
Method | AP | mAP |
---|
Aero | Bike | Bird | Boat | Bottle | Bus | Car | Cat | Chair | Cow | Table | Dog | Horse | Mbike | Person | Plant | Sheep | Sofa | Train | Tv |
---|
YOLO-v3tiny | 65.3 | 70 | 43.8 | 47 | 24.9 | 68.9 | 74.7 | 65.7 | 33.4 | 53.7 | 49 | 61 | 75.3 | 72.1 | 69.1 | 26.9 | 59 | 50.9 | 75 | 60.8 | 57.3 | DS-YOLO | 78.3 | 78.9 | 60.6 | 62.8 | 51 | 76.3 | 85.5 | 76.7 | 49.4 | 70.8 | 66.1 | 71.3 | 80 | 79.9 | 80.9 | 36.9 | 66.8 | 66.2 | 81.7 | 68 | 69.4 | YOLO-v3 | 84.5 | 85 | 77.2 | 68.8 | 65.4 | 85.2 | 86.3 | 86.3 | 62.3 | 79.9 | 74.8 | 85.7 | 87.4 | 86.2 | 81.1 | 50.8 | 80.2 | 77.4 | 82.7 | 75.6 | 78.1 |
|
Table 4. Comparison of DS-YOLO and YOLO%
Method (size) | Year | Base network | mAP /% | Speed /(frame/s) |
---|
Faster RCNN | 2015 | VGG16 | 73.2 | 7 | SSD(300)SSD(300)SSDLite(300)DSSD(321)STDN(300) | 20162017201720172018 | VGG16MobileNetv2MobileNetResNet-101DenseNet-169 | 74.377.472.778.678.1 | 4650569.541.5 | YOLOv2(416)YOLOv3tiny(416)YOLOv3(416) | 201720182018 | Darknet19-Darknet53 | 76.857.178.3 | 6716830 | DS-YOLO | 2019 | - | 69.4 | 141 |
|
Table 5. Comparison of DS-YOLO and others algorithms in VOC2007, 2012
Method | Weight /MB | BFLOPS |
---|
YOLOv3tinyYOLOv3 | 34236 | 5.5765.86 | DS-YOLO | 25 | 8.58 |
|
Table 6. Comparison of DS-YOLO and YOLO
Method | AP, IoU | AP, Area |
---|
0.5∶0.95 | 0.5 | 0.75 | S | M | L |
---|
YOLOv3tinySSD | 14.423.2 | 33.141.2 | 16.323.4 | 3.25.3 | 18.523.2 | 30.239.6 | DS-YOLOYOLOv3 | 20.431.1 | 40.355.3 | 18.135.2 | 4.814.2 | 21.334.1 | 34.746.4 |
|
Table 7. Comparison of DS-YOLO and YOLO、SSD in COCO