基于双目相机和深度学习的鱼类摄食强度分析方法

Design and Experimentation of Fish Feeding Intensity System Based on Binocular Cameras and Deep Learning

摘要:

为精确判别深海网箱养殖中鱼类摄食强度，实现精量投喂，以金鲳鱼进食时造成的水花为研究对象，利用双目相机拍摄到的深度图像进行非侵入性的摄食强度分析，提出一种基于双目相机和深度学习的水花面积语义分割和计算方法。首先，为了使模型能够在低成本的边缘设备上部署，通过StarNet和BiFPN以及自主设计的SCD-Head共享卷积检测头对YOLO v8n-seg进行改进，提出轻量化的YOLO v8n-SBS模型。在精度提升3.2个百分点的同时，参数量与浮点运算量分别减少71%和36%。其次，为降低设备成本，采用双目相机，基于深度信息利用线性回归提出水花面积计算模型DI。最终，两个模型结合为YOLO v8n-SBS-DI，该模型能够对水花进行分割并计算面积，以便通过水花面积变化趋势评估摄食强度。海上试验计算结果显示，水花面积R2为0.914，RMSE为0.973 m2，MAE为0.870 m2。试验结果表明，该模型具有较强鲁棒性，满足复杂环境下水花面积计算需求，可为判别鱼类摄食强度提供技术支持。

Abstract:

Assessing the feeding intensity of fish in large-scale cages is crucial for enhancing feed utilization and reducing farming costs. Traditional feeding methods heavily rely on the experience of aquaculture managers, often leading to overfeeding, which contaminates water quality, or underfeeding, and adversely affects fish health. To accurately determine fish feeding intensity in deep-sea cage farming and achieve precise feeding, focusing on the splashes generated by pompano during feeding, utilizing depth images captured by a binocular camera, a non-invasive feeding intensity analysis method was proposed, involving semantic segmentation and area calculation of the splash. Firstly, to enable the model’s deployment on low-cost edge devices, the YOLO v8n-seg model was improved through the incorporation of StarNet, BiFPN, and a custom-designed SCD-Head shared convolutional detection head, resulting in the lightweight YOLO v8n-SBS model. This modification achieved a 3.2 percentage points increase in accuracy while reducing the number of parameters and floating-point operations by 71% and 36%, respectively. Secondly, to minimize equipment costs, a binocular camera was employed, and PVC boards were used to simulate splash targets on land for experimental convenience. A linear regression model (DI) was proposed to calculate splash area-based on depth information. The results of the DI model on the test set demonstrated an R2 value of 0.977, an RMSE of 0.033 m2, and an MAE of 0.023 m2, indicating robust performance. Ultimately, the two models were combined into YOLO v8n-SBS-DI, which can segment splashes and compute their area, allowing for the assessment of feeding intensity through the trend of splash area changes. Sea trial results showed that the calculated splash area yields an R2 value of 0.914, an RMSE of 0.973 m2, and an MAE of 0.870 m2. These experimental outcomes confirmed that the model exhibited strong robustness and met the demands for splash area calculations in complex environments, thereby providing technical support for determining fish feeding intensity.