Underwater optical image feature extraction based on lightweight SuperPoint network

In view of the poor robustness of feature extraction in underwater vision tasks such as image registration and 3D reconstruction caused by the decline in underwater optical image quality, an lightweight SuperPoint network was proposed. This network addressed the common challenges of detail degradation in underwater optical images, including color distortion and blurring. By leveraging an attention mechanism, it constructed a frequency-spatial dynamic attention fusion module that integrated feature information from both the frequency and spatial domains, thereby enhancing the network’s capability for feature extraction in underwater degraded images. A residual feature enhancement depthwise separable convolutional module was constructed to reduce model complexity and enhance the feature extraction ability of the network. Verification results demonstrate that, compared with the SuperPoint network, the network proposed in this paper achieves a 13.8% reduction in the number of parameters, an 8.0% decrease in computational complexity, and a 31.7% improvement in frame rate. Meanwhile, its repeatability rates under illumination variation and viewpoint variation are increased by 2.3% and 2.1%, respectively. In addition, the network exhibits excellent robustness in feature extraction in the performance evaluation of feature point detection and matching on the SQUID and FLSea datasets.