Estimation of Injection Site in Turbot Based on Image Segmentation

Abstract

<b><sc>Abstract.</sc></b> Fish vaccination plays a vital role in the prevention of fish diseases. Inappropriate injection site will cause low immunization rate even death. Currently, traditional visual algorithms have poor robustness and low accuracy due to the specificity of turbot fins in the application of automatic vaccination systems. To address this problem, we propose a new method for estimating the injection site of turbot based on image segmentation. Many kinds of Deep Learning Networks were used for extraction of background, fish body, side fin, caudal fin. In the subsequent step, the segmentations obtained from the best network were used for calculating body area, body length and body width. These variables were used as covariates in linear models for estimation of the injection site. A network named DeeplabV3+ achieved the best results for intersection over union on the test set of 99, 94, 83 and 90 percent for background, fish body, side fin, caudal fin, respectively. The overall best predictive model achieved R² of 0.92 and 0.95 for the horizontal and vertical coordinates of the injection site, which are almost within the injectable range. In conclusion, the devised method was able to correctly differentiate fish body from background and fins, while the extracted area of fish could be successfully used for estimation of the injection site.