Work

Research on detection technology of biofouling organisms on marine aquaculture cages based on image enhancement algorithms and improved Faster R-CNN detection algorithms

Year: 2025

Type: article

Abstract: • Improved detection accuracy The enhanced Faster R-CNN model demonstrated an average precision of 94.27%, a substantial enhancement over the original model and other prevalent detection models, including YOLOv8 and SSD512. • Enhanced image processing The proposed algorithm, designated as GWA-MSRCR, is predicated on the grayscale world assumption and MSRCR. In comparison with conventional methodologies, the proposed algorithm enhances the peak signal-to-noise ratio (PSNR) of underwater images by 2.47 decibels (dB) and elevates the average precision by 11.27%. • Optimize the efficiency of the model The training weight of the enhanced model is reduced to approximately one-fifth of that of the original Faster R-CNN, thereby enhancing the real-time detection of marine fouling organisms. • Optimization of IoU threshold Research findings indicate that the optimal balance between precision and recall is achieved at an IoU threshold of 0.4, yielding an F1 score of 86.71% and an average precision of 94.27%. • Practical application in aquaculture The proposed method has demonstrated efficacy in the detection and identification of biofouling organisms affixed to marine aquaculture cages. This method provides a practical solution for the automated and intelligent cleaning of marine nets. • Powerful feature extraction The incorporation of the ResNeXt50 backbone and the CBAM module has been demonstrated to enhance the model's capacity to navigate intricate underwater environments and facilitate the detection of diminutive targets. This integration contributes to an augmentation in the model's overall robustness. Traditional marine aquaculture cages produce large amounts of marine fouling organisms that are difficult and labor-intensive to clean. There is an urgent need for drones to replace manual cleaning of underwater cages. One of the key technologies for the automation and intelligent cleaning of drones is the detection and identification of biofouling organisms on marine aquaculture cages. Therefore, this study proposes a technical method for detecting and identifying biofouling organisms on marine aquaculture cages. This method is based on an improved Faster R-CNN detection model. The ResNeXt50 network replaces the original VGG16 feature extraction network. The CBAM module has been integrated into the main network to enhance the model's ability to extract target information. Simultaneously, the proportion and size of the anchor frames are adjusted according to the size of the biofouling organisms, and an appropriate IoU threshold is selected. We propose an image enhancement algorithm for biofouling organisms on marine aquaculture cages. This algorithm is based on the grayscale world assumption and the MSRCR algorithm. To prove the effectiveness of the proposed technique, we conducted a series of comparative experiments. The results show that the improved Faster R-CNN detection model has an average precision of 94.27%, which is 10.31 percentage points higher than the original model, while the training weights are approximately one-fifth the original size. At an IoU threshold of 0.5, the mean average precision (mAP) value is 93.14%, outperforming YOLOv8, YOLOv5, YOLOv11, SSD512, HTDet, Fast R-CNN and Mask R-CNN by 25.37%, 28.86%, 22.78%,13.51%,2.8%,12.6% and 1.17% respectively. Using the image-enhanced dataset for model training also resulted in an average precision that was 11.72 percentage points higher than when using the original dataset. In summary, the technical methods proposed in this paper accurately and quickly detect and identify biofouling organisms on marine aquaculture cages.

Source: Smart Agricultural Technology

DOI: 10.1016/j.atech.2025.101419

Authors: Pengshuai Zhu, Hao Li, Junhua Chen, Chenjun Guo

Institutions: Ningbo University, College of Science & Technology Ningbo University

Language: English

Topic: Advanced Neural Network Applications

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

SDG: Life below water

Open Access status: gold

Estimated APC paid: $1,770