基于改进动态图卷积神经网络的人工鱼礁识别提取方法研究

Study on Artificial Reef Identification and Extraction Method Based on Improved Dynamic Graph Convolutional Neural Network

摘要:

人工鱼礁是一种人为的用于增强鱼类种群，减缓自然栖息地退化，并影响海洋生物资源发展的人工设施。对投放的人工鱼礁进行相应的监测和评估对保护渔业资源发挥重要作用。基于侧扫声呐的图像数据暴露出三维信息缺失和图像变形等问题，利用多波束点云进行人工鱼礁等目标的识别提取成为新的研究方向。本文改进了动态图卷积神经网络（DGCNN）模型，该方法引入法向量特征提取模块和多头注意力机制模块，融合DGCNN中的边卷积特征和法向量特征进行重点关注。本文选取两种代表性鱼礁类型进行训练和试验，将改进方法与PointNet、PointNet++、DGCNN等经典方法进行结果对比。实验结果表明，改进的DGCNN模型在人工鱼礁提取完整度和正确度上优于传统经典方法，具有快速识别、提取精度更高、可信度更高等优势。

Abstract:

Artificial reefs are manmade facilities designed to enhance fish populations, slow down the degradation of natural habitats, and affect the development of Marine biological resources. The corresponding monitoring and assessment of the artificial reefs that have been deployed play an important role in protecting fishery resources. The image data based on sidescan sonar exposes the loss of threedimensional information and image deformation. Using multibeam point clouds for the recognition and extraction of artificial reefs has become a new research direction. In this paper, an improved DGCNN model method has been put forward. A normal vector feature extraction module and a multihead attention mechanism module have been introduced, and the edge convolution features and normal vector features in DGCNN for focused attention have been fused. In this paper, two representative types of fish reefs have been selected for training and testing. The improved method has been compared with PointNet, PointNet++ and DGCNN in terms of results. The experimental results show that the improved DGCNN model outperforms traditional classic methods in the completeness and accuracy of artificial fish reef extraction，and has the advantages of rapid recognition, higher extraction accuracy and higher credibility.