.png){kind=icon}
.png){kind=icon}
Return
Non-intrusive mass estimation method for crucian carp using instance segmentation and point cloud processing
.png){kind=icon}
PRE
AI
DOI：10.1016/j.compag.2024.109445.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Estimating fish mass non-intrusively is critical for precision feeding, stocking density control and optimal fishing times determination in aquaculture. Despite this, challenges such as occlusion, bending, tail swinging and inappropriate imaging angles hinder fully automated mass measurement. Aiming at the above problem, this paper proposes a non-intrusive mass estimation method based on instance segmentation and point cloud processing. Firstly, an instance segmentation model employing YOLOv8-CGBlock-BiFPN was used to delineate fish body contours. Secondly, an automated method for extracting the feature values of the fish based on three-dimensional point clouds was developed. This method, integrating fish contours with stereo vision, applied principal component analysis (PCA) to correct fish orientation before extracting feature values. Finally, the fish mass was estimated using the eXtreme Gradient Boosting (XGBoost) algorithm. Crucian carp was taken as the experimental object and the proposed fish mass estimation method was validated on a real dataset, achieving a mean absolute error (MAE) of 0.01236, a root mean square error (RMSE) of 0.01597, a mean absolute percentage error (MAPE) of 4.51 %, and a coefficient of determination (R-2) of 0.9677. Compared with other methods including Support Vector Regression (SVR)-Linear, SVR-Poly, SVR-Rbf, Linear Regression, Long Short-Term Memory (LSTM), and Back-Propagation Neural Network (BPNN), this approach showed improved performance across all evaluation metrics. The results demonstrate that the proposed method can accurately and non-intrusively estimate the mass of underwater free-swimming fish.
Keywords:
Fish mass estimation
Instance segmentation
Three-dimensional point cloud
Stereo vision
Computer vision
Journal
IF:
8.9
Papers:
9.9K
Citations:
4.8W