Deep Learning Approach For Objects Detection in Underwater Pipeline Images

Figures & data

Figure 1. Yolov1 architecture.

Figure 2. The detection pipeline of YOLO.

Figure 3. Yolov4 structure.

Figure 4. Yolov4 tiny network structure.

Figure 5. An overview of object detection with faster R-CNN.

Figure 6. Snapshot of labelling tool.

Table 1. Dataset description.

Dataset	Description
Annotated Image:	3021 annotations
Number of Classes	5
Training	2415 images
Validation	303 images
Testing	303 images

Table 2. Training parameters configuration.

Model	Batch size	Subdivision	Width x Height	Momentum	Decay	Learning rate	Activation
YOLO^1	64	64	416x416	0.949	0.00005	0.0013	Mish
Tiny^2	64	64,16	416x416	0.9	0.0005	0.00261	Leaky
CSP^3	64	64	416x416	0.949	0.0005	0.001	Mish
Mod^4	64	32	416x416	0.9	0.0005	0.0001	Leaky
Mod^5	64	8	416x416	0.9	0.0005	0.0001	Leaky

Figure 7. Visual representation of IoU criterion.

Figure 8. Loss function and mAP performance.

Table 3. Performance evaluation result.

Architecture	Accuracy					mAP(%)
	pipeline	leakage point	concrete weight	concrete mat	pipe coupling
YOLOv4	98.53	81.83	94.62	96.00	96.40	94.21
YOLOv4-Tiny@64	54.03	66.78	76.07	83.09	79.39	72.97
YOLOv4-Tiny@16	86.84	81.27	96.05	98.78	97.46	92.43
CSP-YOLOv4	91.04	84.37	97.85	98.80	85.03	93.97
YOLOv4 @ResNet50	78.05	34.92	84.70	95.48	89.38	79.50
YOLOv4 @DenseNet201	91.26	63.83	90.69	97.92	95.65	88.40
Faster RCNN	65.48	45.99	72.37	80.45	76.10	70.49

Figure 9. Detection performance of YOLO models.

Table 4. Processing speed inference.

Architecture	Frame Per Second(FPS)
YOLOv4	25
YOLOv4-Tiny	38
CSP YOLOv4	15
YOLOv4@ResNet50	17
YOLOv4@DenseNet201	16