Fishery Detection and Counting Model uing Fish Images

Abstract Fishing in Senegal is threatened by overfishing, compounded by a lack of fisheries management. Stock replenishment and fish classification are generally done manually, and catches are not always declared. In addition, the collection of fishing data suffers from a lack of tools for monitoring and counting the fish caught at the wharves. Although researchers have carried out studies on the fishery in Senegal, data collection is virtually non-existent, and there is no local database dedicated to the fishery or automatic detection and counting algorithm. In this article, a model for automatic detection and counting of fished species is proposed, using a semantic segmentation algorithm. The data used to form the adapted local database are collected from fish images taken at the Soumbédioune fishing wharf in Senegal. This database is supplemented by the Fishbase. This collected data is then segmented to form the appropriate local database. This database is used in conjunction with YOLO v8, an essential element in the detection of images with bounding boxes, to train the model. The results obtained are very promising for the proposed automatic fish detection and counting model. For example, the recall-confidence scores reflect the performance of bounding boxes, with scores ranging from 0.01 to 0.75, thus confirming the effectiveness of the model with bounding boxes. These results are of great importance for improving fisheries policies in Senegal.