This paper proposes a machine vision based method for quantifying fish feeding behavior. By integrating fish movement features and image texture features, the method can achieve efficient and accurate quantification. The study introduces an improved MC-YOLO object detection model, which enhances feature extraction capabilities through the Multi-Scale Lightweight Convolution (MSLConv) module and the Convolutional Block Attention Module (CBAM), while maintaining low computational complexity. Experiments conducted in an industrial Recirculating Aquaculture System (RAS) demonstrate that this method has significant advantages in detection accuracy, feeding state recognition, and adaptability to complex breeding environments. It provides strong technical support for precise feeding and has broad application prospects and promotion value.