1 Introduction
2 Data set
2.1 Data collection
2.2 Data cleaning
3 Methods and experiment
3.1 The YOLO models
3.2 Structure of YOLOv4 model
3.3 The loss function
3.4 Experiments setup
3.5 Evaluation criteria
4 Results and analysis
4.1 Evaluation of three models
Table 1 Comparison results of different models with parameters score = 0.35 and IOU = 0.5 |
Models | Precision/% | Recall/% | F1-score/% | mAP |
---|---|---|---|---|
Faster-RCNN | 85.34 | 75.66 | 80.85 | 76.00 |
YOLOv2 | 77.00 | 73.00 | 75.00 | 71.52 |
YOLOv3 | 86.00 | 77.00 | 81.00 | 76.96 |
YOLOv4 | 87.00 | 79.00 | 83.00 | 78.99 |
4.2 Evaluation of different IOU thresholds
Table 2 The impact of IOU values on the performance of different models |
model | IOU | Precision/% | Recall/% | F1/-score% | mAP/% |
---|---|---|---|---|---|
YOLOv2 | 0.20 | 85.00 | 81.00 | 83.00 | 84.05 |
0.35 | 83.00 | 80.00 | 81.00 | 81.03 | |
0.50 | 77.00 | 73.00 | 75.00 | 71.52 | |
0.65 | 53.00 | 51.00 | 52.00 | 40.64 | |
YOLOv3 | 0.20 | 91.00 | 81.00 | 86.00 | 84.05 |
0.35 | 90.00 | 81.00 | 85.00 | 82.36 | |
0.50 | 86.00 | 77.00 | 81.00 | 76.96 | |
0.65 | 65.00 | 59.00 | 62.00 | 48.55 | |
YOLOv4 | 0.20 | 92.00 | 84.00 | 87.00 | 85.01 |
0.35 | 91.00 | 83.00 | 87.00 | 83.69 | |
0.50 | 87.00 | 79.00 | 83.00 | 78.99 | |
0.65 | 70.00 | 64.00 | 67.00 | 56.38 |
4.3 Evaluation of the foxtail millet ear detection with/without anchor boxes adjustment
Table 3 Comparison results of models with/without anchor boxes adjustment |
Model | Precision/% | Recall/% | F1-score/% | mAP |
---|---|---|---|---|
YOLOv3 | 86.00 | 77.00 | 81.00 | 76.90 |
YOLOv3_adj | 87.00 | 78.00 | 81.00 | 77.16 |
YOLOv4 | 87.00 | 79.00 | 83.00 | 78.99 |
YOLOv4_adj | 87.00 | 80.00 | 83.00 | 80.87 |
4.4 Evaluation of the changing reasons of model criteria
Table 4 TP and FP values predicted by the experimental models for the ear target |
Model | TP | FP | TP increment | FP increment | mAP increment |
---|---|---|---|---|---|
YOLOv2 YOLOv3 | 2052 2157 | 623 352 | 0 105 | 0 -271 | 0 5.44 |
YOLOv4 | 2220 | 329 | 168 | -294 | 7.47 |
4.5 Evaluation of the foxtail millet ear detection with different input original image size
Table 5 The impact of different original image size on the foxtail millet ear detection |
Original image size/px | YOLOv4 input size/px | Precision/% | Recall/% | F1-score/% | mAP |
---|---|---|---|---|---|
4864×3648 | 608×608 | 87 | 79.00 | 83.00 | 78.99 |
2000×1500 | 608×608 | 92 | 84.00 | 88.00 | 83.50 |