Underwater target recognition system based on the improved YOLOv5
The underwater target recognition algorithm has emerged as a highly significant research direction within target recognition, with numerous new algorithms being proposed in recent years. Among these, the YOLOv5 algorithm has gained widespread recognition for its exceptional balance between detection...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis-Master by Coursework |
| Language: | English |
| Published: |
Nanyang Technological University
2024
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/179885 |
| _version_ | 1826128232421785600 |
|---|---|
| author | Pan, Yawen |
| author2 | Su Rong |
| author_facet | Su Rong Pan, Yawen |
| author_sort | Pan, Yawen |
| collection | NTU |
| description | The underwater target recognition algorithm has emerged as a highly significant research direction within target recognition, with numerous new algorithms being proposed in recent years. Among these, the YOLOv5 algorithm has gained widespread recognition for its exceptional balance between detection accuracy and speed. However, when dealing with small targets and various complex underwater environments, YOLOv5 still encounters issues with missed and false detections, posing challenges for practical applications. To enhance the recognition performance of YOLOv5, this paper proposes three improvements based on the YOLOv5s model, the lightest variant in the YOLOv5 series. The improved model can accurately identify three categories: plastics, rovs, and bios, so as to judge the surrounding environment of the UUV. The main enhancements are as follows:
Firstly, to improve the accuracy of underwater target localization with the YOLOv5s algorithm, a Selective Kernel (SK) attention mechanism module was integrated into the Backbone output. This attention mechanism module boosts the network's ability to extract channel feature information without altering the number or size of channels. It is also flexible in configuration and usage. This enhanced model is referred to as YOLOv5s_SKA.
Secondly, to enhance the detection capabilities of the YOLOv5s algorithm for underwater targets, a Bi-directional Feature Pyramid Network was integrated into the Neck module. BiFPN functions as an efficient bidirectional feature fusion network, enabling comprehensive fusion of multi-scale features. This enhancement strengthens the ability of model to perceive small-scale underwater targets, thereby improving recognition accuracy. This upgraded model is referred to as YOLOv5s_BiFPN.
Thirdly, the KL loss function is adopted to reduce the impact of ambiguous bounding boxes on the prediction results and improve the recognition rate of occluded underwater targets. This improved model is called YOLOv5s_KL LOSS.
Finally, this study utilizes underwater datasets to conduct comparative experiments with the original YOLOv5s algorithm. Results indicate that YOLOv5s SKA demonstrates a 1.2% improvement in mAP compared to the baseline, achieving 79.8%. YOLOv5s BiFPN shows an increase of 0.8% in mAP, reaching 79.4%. YOLOv5s_KL loss achieves a 0.4% mAP increase, reaching 79.0%. YOLOv5s_Mix, integrating three enhancement strategies, achieves the highest increase of 2.2% in mAP, reaching 80.8%. In conclusion, the proposed enhancements effectively enhance the algorithm's detection accuracy across the board. The improved model can accurately identify three categories: plastics, rovs, and bios. |
| first_indexed | 2024-10-01T07:21:34Z |
| format | Thesis-Master by Coursework |
| id | ntu-10356/179885 |
| institution | Nanyang Technological University |
| language | English |
| last_indexed | 2024-10-01T07:21:34Z |
| publishDate | 2024 |
| publisher | Nanyang Technological University |
| record_format | dspace |
| spelling | ntu-10356/1798852024-09-06T15:43:49Z Underwater target recognition system based on the improved YOLOv5 Pan, Yawen Su Rong School of Electrical and Electronic Engineering RSu@ntu.edu.sg Engineering Underwater target Attention mechanism Bi-directional feature fusion Loss function The underwater target recognition algorithm has emerged as a highly significant research direction within target recognition, with numerous new algorithms being proposed in recent years. Among these, the YOLOv5 algorithm has gained widespread recognition for its exceptional balance between detection accuracy and speed. However, when dealing with small targets and various complex underwater environments, YOLOv5 still encounters issues with missed and false detections, posing challenges for practical applications. To enhance the recognition performance of YOLOv5, this paper proposes three improvements based on the YOLOv5s model, the lightest variant in the YOLOv5 series. The improved model can accurately identify three categories: plastics, rovs, and bios, so as to judge the surrounding environment of the UUV. The main enhancements are as follows: Firstly, to improve the accuracy of underwater target localization with the YOLOv5s algorithm, a Selective Kernel (SK) attention mechanism module was integrated into the Backbone output. This attention mechanism module boosts the network's ability to extract channel feature information without altering the number or size of channels. It is also flexible in configuration and usage. This enhanced model is referred to as YOLOv5s_SKA. Secondly, to enhance the detection capabilities of the YOLOv5s algorithm for underwater targets, a Bi-directional Feature Pyramid Network was integrated into the Neck module. BiFPN functions as an efficient bidirectional feature fusion network, enabling comprehensive fusion of multi-scale features. This enhancement strengthens the ability of model to perceive small-scale underwater targets, thereby improving recognition accuracy. This upgraded model is referred to as YOLOv5s_BiFPN. Thirdly, the KL loss function is adopted to reduce the impact of ambiguous bounding boxes on the prediction results and improve the recognition rate of occluded underwater targets. This improved model is called YOLOv5s_KL LOSS. Finally, this study utilizes underwater datasets to conduct comparative experiments with the original YOLOv5s algorithm. Results indicate that YOLOv5s SKA demonstrates a 1.2% improvement in mAP compared to the baseline, achieving 79.8%. YOLOv5s BiFPN shows an increase of 0.8% in mAP, reaching 79.4%. YOLOv5s_KL loss achieves a 0.4% mAP increase, reaching 79.0%. YOLOv5s_Mix, integrating three enhancement strategies, achieves the highest increase of 2.2% in mAP, reaching 80.8%. In conclusion, the proposed enhancements effectively enhance the algorithm's detection accuracy across the board. The improved model can accurately identify three categories: plastics, rovs, and bios. Master's degree 2024-09-02T06:15:23Z 2024-09-02T06:15:23Z 2024 Thesis-Master by Coursework Pan, Y. (2024). Underwater target recognition system based on the improved YOLOv5. Master's thesis, Nanyang Technological University, Singapore. https://hdl.handle.net/10356/179885 https://hdl.handle.net/10356/179885 en ISM-DISS-03768 application/pdf Nanyang Technological University |
| spellingShingle | Engineering Underwater target Attention mechanism Bi-directional feature fusion Loss function Pan, Yawen Underwater target recognition system based on the improved YOLOv5 |
| title | Underwater target recognition system based on the improved YOLOv5 |
| title_full | Underwater target recognition system based on the improved YOLOv5 |
| title_fullStr | Underwater target recognition system based on the improved YOLOv5 |
| title_full_unstemmed | Underwater target recognition system based on the improved YOLOv5 |
| title_short | Underwater target recognition system based on the improved YOLOv5 |
| title_sort | underwater target recognition system based on the improved yolov5 |
| topic | Engineering Underwater target Attention mechanism Bi-directional feature fusion Loss function |
| url | https://hdl.handle.net/10356/179885 |
| work_keys_str_mv | AT panyawen underwatertargetrecognitionsystembasedontheimprovedyolov5 |