Deep-sea pollutant detection for autonomous underwater robots

Fig. 2. Predicted values of grid cells for bounding boxes and categories

Fig. 3. Comparison of feature fusion networks

Fig. 4. Structure of adown module

Fig. 5. Structure of efficient head

Fig. 6. Structure of diversified branching block

Fig. 7. Calculation factor for bounding box regression

Fig. 8. Selected images from Debris dataset

Fig. 9. Structure of underwater image restoration algorithm

Fig. 10. Restoration result of underwater image by different algorithms

Fig. 11. Training results for 300 epochs

Fig. 12. Confusion matrix for Debris-yolo model

Fig. 13. Feature images of Debris-yolo model

Fig. 14. Visualisation of detection results from different model sections

Table 1. Results of ablation experiments

Table 1. Results of ablation experiments

Model	A	B	C	P	R	mAP_0.5	mAP_0.5∶0.95	Parameter	GFLOPS
0	-	-	-	88.6	75.2	83.8	62.9	3.15	8.20
1	√	-	-	90	76.2	85.0	63.2	1.83	6.90
2	-	√	-	88.3	75.9	84.7	63.8	3.66	8.10
3	-	-	√	88.2	78.0	84.9	64.0	2.88	8.20
4	√	√	-	88.3	76.9	84.6	63.1	1.63	5.60
5	√	-	√	87.9	77.3	84.8	62.9	1.83	6.90
6	-	√	√	88.9	77.6	84.9	64.5	3.66	8.10
7	√	√	√	89.0	78.2	85.1	64.5	1.63	5.60

Table 2. Comparison of application data preprocessing effects
View table
View in Article
Table 2. Comparison of application data preprocessing effects
Model P R mAP_0.5 mAP_0.5∶0.95
0 86.2 72.2 80.7 61.7
1 89.0 78.2 85.1 66.4

Table 3. Comparison of detection results of different models
View table
View in Article
Table 3. Comparison of detection results of different models
Model P R mAP_0.5 mAP_0.5：0.95
ResNet50-Faster R-CNN 82.5 72.3 78.9 58.2
YOLOv5s 88.2 76.8 83.5 60.9
DSDebrisNet 88.3 77.0 83.9 61.4
YOLOv8n 88.6 75.2 83.8 62.9
YOLOv10n 87.7 77.1 83.6 63.8
Debris-yolo 89.0 78.2 85.1 64.5

Table 4. Comparison of results of different models for each category of detection

Table 4. Comparison of results of different models for each category of detection

Modal	Rov	Eel	Fish	Fabric	Fishing_net_Rope	Glass	Metal	Natural_debris	Plastic	Rubber
ResNet50-Faster R-CNN	71.6	84.6	85.7	94.2	73.6	93.6	80.7	79.9	72.5	91.5
YOLOv5s	75.5	86.5	86.2	93.6	72.2	92.1	81.5	81.6	72.3	93.2
DSDebrisNet	74.7	85.5	86.4	93.4	71.8	93.8	83.5	79.9	73.3	96.7
YOLOv8n	74.6	84.4	87.6	94.6	72.5	95.1	80.7	81.6	73.3	93.7
YOLOv10n	70.7	90.2	87.8	93.9	72.6	94.4	82.0	81.0	74.0	89.8
Debris-yolo	78.3	87.6	88.3	96.0	73.7	92.4	83.9	82.1	76.5	91.8

Table 5. Comparison of different model lightweight metrics

Table 5. Comparison of different model lightweight metrics

Model	Detection speed/（frame·s^-1）	Parameters/MB	GFLOPS	Size/MB
Faster R-CNN	19.11	41.39	208.00	478
YOLOv5s	102.04	6.71	15.80	14.5
DSDebrisNet	95.24	6.71	15.80	13.8
YOLOv8n	88.49	3.15	8.70	6.00
YOLOv10n	98.45	2.30	66.70	5.50
Debris-yolo	90.91	1.63	5.60	4.33

Table 6. Detection speed and inspection time in different situations