Deep Learning–Based Sturgeon Counting and Length Estimation Using Segmentation and Skeletonization

Abstract

The article discusses the problem of contactless monitoring of sturgeon in aquaculture based on video data. The relevance of the work is related to the fact that digitalization of aquaculture is becoming one of the key areas of the industry's development: farms need to promptly and regularly obtain objective indicators on the condition and size composition of fish without laborious manual procedures. In many practical scenarios, selective trapping, measuring and weighing are still used, which increases time costs and can negatively affect fish. Therefore, solutions are needed to automatically obtain quantitative characteristics from video. An end-to-end method is proposed that combines neural network detection and instance segmentation of sturgeons with subsequent calculation of biometric metrics. According to the predicted masks, the frame is resized and morphological cleaning is performed, then the central line (skeleton) of the object and its length in pixels are calculated. . The result of the algorithm is the indicators for the frame of the video stream: the number of fish detected, the lengths of individual individuals, as well as the average, minimum and maximum lengths. Examples of the method's operation in real-world shooting conditions and the results of training the segmentation model are given