Automatic segmentation of fish using deep learning with application to fish size measurement

One of the leading causes of overfishing is the catch of unwanted fish and marine life in commercial fishing gears. Echosounders are nowadays routinely used to detect fish schools and make qualitative estimates of the amount of fish and species present. However, the problem of estimating sizes using acoustic systems is still largely unsolved, with only a few attempts at real-time operation and only at demonstration level. This paper proposes a novel image-based method for individual fish detection, targeted at drastically reducing catches of undersized fish in commercial trawling. The proposal is based on the processing of stereo images acquired by the Deep Vision imaging system, directly placed in the trawl. The images are pre-processed to correct for nonlinearities of the camera response. Then, a Mask R-CNN architecture is used to localize and segment each individual fish in the images. This segmentation is subsequently refined using local gradients to obtain an accurate estimate of the boundary of every fish. Testing was conducted with two representative datasets, containing in excess of 2600 manually annotated individual fish, and acquired using distinct artificial illumination setups. A distinctive advantage of this proposal is the ability to successfully deal with cluttered images containing overlapping fish.