Abstract: To achieve precise and adaptive feeding control in recirculating aquaculture systems (RAS) and solve problems such as low feed utilisation and coarse growth regulation caused by static feeding strategies, this study proposes an intelligent feeding method integrating visual perception with Deep Q-Network (DQN). Taking freshwater grouper (Cichlasoma managuense) as the subject, real-time tracking of fish movement velocity is achieved through YOLOv8 and DeepSORT, and a quantitative metric of fish school feeding intensity is constructed combined with texture features extracted from the gray-level co-occurrence matrix; This metric, along with water temperature, dissolved oxygen levels, and expected feeding intensity, formed the state input for a multi-objective reward function. A decision model was then trained using DQN, enabling a closed-loop control system. Experimental results showed an average mAP@.5 of 85.3%. With an average total feed input of just 378.4 g per fish, the model achieved a weight gain rate (WGR) of 54.38% and a feed conversion ratio (FCR) of 1.09, both significantly better than those obtained with conventional feeding methods. This approach enables real-time monitoring of feeding behaviour and dynamic adjustment of feeding strategies, offering a robust technical solution for precision management in RAS.