Shrimp length estimation for an automatic feeding-tray lifting system used in shrimp farming

The shrimp farmer monitors shrimp growth by measuring the shrimp’s size with some equipment or estimating its size with the naked eye. These traditional approaches are time-consuming, experience-required, and tedious. To facilitate shrimp farmers, this research aims to propose shrimp length estimation based on images captured by an automatic feeding-tray lifting system used in shrimp farming. The framework of the proposed method was designed into three main processes. The first process is shrimp detection. Two convolutional neural network models for instance segmentation were proposed to detect shrimps and generate segmentation masks. The first model is a mask region-based convolutional neural network with a ResNeXt model which achieved a precision of 97.10%, a recall of 94.36%, a F1 score of 95.71%, and an average precision of 95.97%. The second one is the newest version of the You Only Look Once model which achieved a precision of 91.87%, a recall of 97.74%, a F1 score of 94.72%, and an average precision of 95.60%. From the comparison of both models, the first model was chosen as a primary detection model for the proposed method because of its performance and ability to generate a high-quality segmentation mask. In the second process, this study proposed an approach for determining a reference scale which is used to convert a pixel distance in the image to an actual length. The diameter of a feeding tray installed on the automatic feeding-tray lifting system was used as a reference object. The last process is shrimp length estimation. Two unique methods were proposed for estimating shrimp length. The first method achieved a mean absolute error of 1.63 cm and a mean absolute percentage error of 11.59%. The second method achieved a mean absolute error of 0.65 cm and a mean absolute percentage error of 4.61%. Both methods have their own strength and weakness. The former works better in the case of straight-body shrimp, but the latter works well in the case of curved-body shrimp. To improve the performance, another method was proposed to combine both methods using a box ratio or eigenvalue ratio as a method-selection parameter. The combination method based on the box and eigenvalue ratio achieved a mean absolute error of 0.59 cm and 0.60 cm, respectively. It indicates that the combination method can reduce the error of the previously proposed estimation methods. This proposed method helps shrimp farmers save time and monitor their shrimp growth easily. It also generates useful digital data for shrimp farmers so that they can use it to analyze and manage their farms effectively. Moreover, some features were included in the proposed method in order to facilitate the shrimp farmer, for example, sending an estimation result to the shrimp farmer via a LINE application, notifying problems on the automatic feeding tray lifting system to the shrimp farmer, etc.

虾农通过使用设备测量虾的尺寸或肉眼估算其大小来监测虾的生长情况。这些传统方法耗时、依赖经验且繁琐。为了方便虾农，本研究旨在提出一种基于虾养殖中自动投食盘升降系统拍摄图像的虾体长估算方法。所提方法的框架设计为三个主要流程。第一个流程是虾体检测。研究提出了两种用于实例分割的卷积神经网络模型，以检测虾体并生成分割掩码。第一种模型是结合ResNeXt模型的掩码区域卷积神经网络，其精确率达97.10%、召回率94.36%、F1分数95.71%、平均精度95.97%。第二种模型是You Only Look Once的最新版本，其精确率为91.87%、召回率97.74%、F1分数94.72%、平均精度95.60%。通过对比两种模型，由于其性能及生成高质量分割掩码的能力，第一种模型被选为所提方法的主要检测模型。第二个流程中，本研究提出了一种确定参考尺度的方法，用于将图像中的像素距离转换为实际长度。自动投食盘升降系统上安装的投食盘直径被用作参考对象。最后一个流程是虾体长估算。研究提出了两种独特的虾体长估算方法。第一种方法的平均绝对误差为1.63厘米，平均绝对百分比误差为11.59%；第二种方法的平均绝对误差为0.65厘米，平均绝对百分比误差为4.61%。两种方法各有优劣：前者在虾体伸直时表现更佳，而后者适用于虾体弯曲的情况。为提升性能，研究提出了一种结合两种方法的策略，以框比例或特征值比例作为方法选择参数。基于框比例和特征值比例的组合方法分别取得了0.59厘米和0.60厘米的平均绝对误差，表明该组合方法可降低先前提出的估算方法的误差。该方法可帮助虾农节省时间并轻松监测虾的生长情况，还能为虾农生成有用的数字数据，以便其有效分析和管理养殖场。此外，所提方法还包含一些便利虾农的功能，例如通过LINE应用向虾农发送估算结果、向虾农通知自动投食盘升降系统的故障等。