Abstract
The complex optical environment underwater leads to spectral selective absorption and scattering effects that significantly degrade image quality, posing challenges for fish target recognition. Additionally, current detection models face limitations in real-time applications due to their high computational complexity. To address these issues, this study introduces a two-stage fusion enhancement strategy for an underwater fish intelligent recognition framework. In image pre-processing, a collaborative enhancement mechanism combining CLAHE and Auto Levels is developed to enhance contrast while preserving image naturalness. A local adaptive processing model is established, and a dynamic color gamut compensation algorithm is integrated to correct color distortion effectively. The introduction of the ADown module optimizes the feature extraction network, reducing computational complexity. Furthermore, a VoV-GSCSP feature multiplexing mechanism is devised to enhance feature acquisition ability while maintaining precision. Experimental findings demonstrate that the model enhances detection precision and mAP@50 by 6.8% and 3.0%, respectively. Moreover, model volume and parameters are reduced by 12.7% and 11.5%. The F1 score is improved by 0.03, surpassing the baseline model in overall performance. The recognition robustness and adaptability in complex underwater environments are effectively enhanced.
DOI: 10.61416/ceai.v27i4.9664