双目视觉下基于NGBoost的鱼体质量估算方法

Fish Mass Estimation Method Based on NGBoost under Binocular Vision

摘要:

鱼体质量对于评判鱼类生长状况、促进精准投喂和提高水产养殖效益至关重要。为实现精准无损的鱼体质量估算，本文提出一种基于双目相机的双维度特征提取和自然梯度提升 （Natural gradient boosting，NGBoost） 方法。首先通过双目相机获取鱼体图像，并进行相机标定和图像校正操作;其次利用图像处理技术对校正后的图像分割获得鱼体目标，提取出鱼体目标的二维特征;在此基础上进行立体匹配获得鱼体视差图，提取鱼体左右图像的对应关键匹配点，并利用三角变换原理计算三维空间特征点坐标，实现鱼体目标三维特征的提取;最后采用基于NGBoost的方法预测出鱼体质量。本文不仅提取二维平面特征，还提取体长、体宽和鱼体深度比值三维空间特征，实现了鱼体多维特征的提取，丰富了模型的特征表示，解决了单平面维度特征导致的质量预测不准确问题。本文以鲫鱼为实验对象，在真实数据集上进行实验，结果表明，平均绝对误差为0.006 3 kg，均方根误差为0.008 7 kg，决定系数为0.928 7。此外，与多种质量估算方法进行对比，本文方法的各评价指标均有较大幅度提升，能够较为准确地预测出鱼体质量。

Abstract:

Fish mass is crucial for evaluating fish growth status, promoting precise feeding in aquaculture, and improving aquaculture efficiency. To accurately estimate fish mass, a fish mass estimation method based on dual dimensional feature extraction and natural gradient boosting （NGBoost）was proposed under the premise of using binocular cameras. Firstly, fish images were obtained through a binocular camera, and camera calibration and image correction operations were performed. Secondly，image processing technologies were used to segment the corrected image to obtain the fish target, and the two-dimensional features of the fish target were extracted. On this basis, stereo matching was performed to obtain the fish disparity map, extract the corresponding key matching points of the left and right images of the fish, and calculate the coordinates of the three-dimensional spatial feature points by using the triangular transformation principle, achieving the extraction of the three -dimensional features of the fish target. Finally, the method based on NGBoost was used to predict fish mass. Different dimensional features of fish from two - dimensional plane and three - dimensional space were extracted, solving the problem of inaccurate prediction of fish mass caused by single -plane dimensional features. At the same time, in addition to common three -dimensional features such as length and width, the fish depth ratio was also extracted, enriching the feature representation of the model and improving the accuracy of fish mass prediction. The crucian carp were taken as the experimental object and the proposed method was tested on the real dataset. The results showed that the mean absolute error（MAE）was 0.006 3 kg, the root mean square error （RMSE）was 0.008 7 kg, and the coefficient of determination（R2）was 0.928 7.Compared with various mass estimation methods, the performance of each evaluation metric of the proposed method has been greatly improved, predicting the fish mass more accurately.