Abstract:As one of the visual attributes of fish appearance, body length is a key factor related to the monitoring of fish growth status, regulation of water environment, feeding of bait drugs, quality and safety of fish products and the estimation of economic benefits. However, traditional body length estimation methods involve processes such as capture, anesthesia and manual measurement, which are time-consuming, labor-intensive and low-precision. In addition, it can also cause physiological stress responses and negatively affect the tested fish. With the rapid development of imaging technology, computing power and hardware equipment, non-destructive measurement methods based on machine vision have emerged rapidly, overcoming the limitations of traditional methods in terms of cost and performance. With its advantages of fast, accurate, timely, efficient and repeatable batch detection, it has become a powerful tool for fish body length measurement and plays a positive role in improving the economic benefits of aquaculture. The existing domestic and foreign research literature was summarized and sorted out, and the machine vision-based image acquisition equipment, fish contour extraction algorithms and length measurement methods were systematically analyzed and discussed. High-efficiency image acquisition and high-quality image data were important guarantees for accurate measurement. The advantages, disadvantages and applications of monocular cameras, binocular cameras based on optical imaging were firstly compared and analyzed. Secondly, the extraction of fish body contours from two parts of traditional image processing technology and image segmentation technology based on deep learning was summarized. Then, it was concluded that the underwater fish segmentation method based on deep learning had better robustness and versatility in the complex underwater scene. Using the image acquisition mode as the classification basis, the body length measurement methods based on the 2D mode and the 3D mode were described respectively. From the perspective of manual participation, the measurement methods based on the 3D mode were divided into automation and semi-automation. The semi-automation of stereo intersection methods such as DLT, template matching, and the Haar classifier were summarized. Also, convex hull algorithm, point cloud, and landmark point geometric morphology measurement method based on fully automated three-dimensional measurement methods were listed. However, due to the difficulty of deploying underwater cameras, the complication of underwater scenes, and the sensitiveness of the measured fish body, it was very challenging to apply machine vision technology to the measurement of fish body length widely. At last, the trend of fish body length measurement based on machine vision was proposed. Furthermore, image enhancement was the research focus, and fish contour extraction based on deep learning methods was the key technology. Also, developing length measurements based on 3D mode was the mainstream method and using three-dimensional point cloud data measurement and geometric features to fit contours was a direction. Machine vision combined with technologies such as deep learning, pattern recognition, and environmental perception, became a key method for obtaining fish growth information, which can provide technical support for the refined and intelligent management of aquaculture.
