Analysis of Deep Learning Methods in Video Analytics, Robotic Monitoring, and Aquaculture Management with an Empirical Comparison of Lightweight YOLO Models for Fish Detection

Abstract

This paper presents a review of state-of-the-art deep learning methods for video analytics in aquaculture over the period 2025–2026, complemented by an empirical evaluation of two lightweight YOLO-family neural network models for the fish detection task. Four key research directions are analyzed: fish disease detection, segmentation and phenotyping, tracking and biomass estimation, and the development of lightweight models. Based on contradictions identified between existing studies, an experimental comparison of two architectural configurations - the original YOLOv26 and YOLOv11 integrated with the Reparameterized Multi-Head attention module (RepMHD) - was conducted on the public DeepFish dataset, which contains images of fish in natural tropical marine habitats. The results show that the YOLOv11 + RepMHD configuration achieves higher Precision (0.978 vs. 0.960) and a comparable mAP@0.5 value (0.991 vs. 0.990) with 2.28 million parameters and 5.6 GFLOPs, whereas the original YOLOv26 demonstrates a higher mAP@0.5:0.95 value (0.816 vs. 0.812) with 2.38 million parameters and 5.2 GFLOPs. These findings indicate that the choice between these two strategies should be guided by the specifics of the application - particularly the priority assigned to minimizing false positive versus false negative detections.