A detection alarm optimization method and system based on a visual language multi-modal large model

Abstract

The application discloses a detection alarm optimization method and system based on a visual language multi-modal large model, which comprises the following steps: acquiring a video stream and performing real-time detection through a target detection model to output a detection result; screening and generating preliminary alarm information according to a preset alarm condition; constructing a question template library and dynamically generating specific questions according to the target category and scene type in the alarm information; inputting the alarm image and the question into a visual language multi-modal large model for semantic reasoning to obtain a verification result containing a judgment label and a semantic confidence; and judging whether to trigger a final alarm and an alarm level based on the detection confidence, the verification result and a multi-level threshold decision rule. The application performs semantic-level verification on the preliminary alarm by introducing a visual language large model, effectively identifies and filters false alarms such as similar false detection and scene mismatch false detection, significantly reduces the false alarm rate, reduces the workload of manual review, enhances the scene adaptability, and is compatible with the existing detection system and easy to integrate and deploy.

Images