Intelligent sorting method of fry based on adaptive imaging closed loop of state space modeling

Abstract

This invention discloses an intelligent fish fry sorting method based on state-space modeling and adaptive imaging closed-loop, relating to the field of aquaculture technology. At the acquisition end, an imaging closed loop of "controllable illumination—frequency domain calibration—exposure and gain self-calibration" is constructed. At the computation end, a lightweight front-end of "multi-scale template matching ROI rapid localization + small-range four-sided approximation tracking" is used. At the detection end, an end-to-end detection network is constructed, with a visual state-space module introduced in the backbone stage. In the feature fusion stage, HSFPN is used, and multi-scale feature selection and complementary fusion are completed through channel attention. At the control end, the detection results are mapped to the coordinate system of the sorting actuator, and combined with trajectory correlation, activity heatmap, abnormal fish fry identification, dead fish fry identification, and closed-loop strategy, the actuator is driven to achieve non-contact hierarchical sorting and counting. This invention, using the above method, improves the detection rate of dense small targets and robustness under complex water imaging conditions while ensuring real-time performance.

Images