Method for underwater transparent target recognitionusing event based cameras

DOI:	10.11809/bqzbgcxb2025.03.025
Keywords:	instance segmentation; underwater transparent organisms; deep learning; event camera; attention mechanism; edge detection; multi feature fusion
Abstract:	A segmentation algorithm for underwater transparent organisms has been proposed, addressing the problem of significantly low visibility and challenging RGB frame image recognition in underwater environments, which is based on event cameras. Firstly, image data was obtained using an event camera, and an underwater transparent organism dataset was constructed, which comprised 6 421 event frames and RGB frames, totaling 12 421 targets. During the encoding stage, the event frames and RGB frames were separately input into a residual network improved with the CBAM attention mechanism to extract features and undergo simple fusion. An Edge Clue Search Module (ECSM) was designed to capture easily overlooked detailed information through simple weighted operations. In the decoding section, a Multi dimensional Feature Fusion Module (MFFM) was designed to more fully integrate various prior information, thereby improving the detection accuracy. Training and testing were conducted on the self made underwater transparent dataset. The algorithm achieved an average precision of 85.8% and a frames per second (FPS) rate of 66.4 f/s, which is an improvement compared to most mainstream methods. This demonstrates that the proposed method not only ensures real time performance in complex underwater environments but also achieves good segmentation results.
Issue:	Vol. 46 No. 3 (2025)
Published:	2025-03-31
PDF