Breaking the Waves: A Region-Based Feature Matching Method in Wave Images Using Improved Coherent Point Drift

The homogeneity and adequacy of keypoint matching are critical to visual measurement of sea waves. However, the sparsely distributed edges and corners in weakly textured scenarios of wave images, can lead to a lack of abundant feature points, which gives rise to an insufficiency of distinct descriptors that can effectively discriminate between different keypoints. In this paper, we proposed a region-Based feature matching method in wave images using improved Coherent Point Drift (CPD). The process begins with segmenting the wave image into superpixels. Landmark-superpixels are then located using a certain keypoint matching method. These superpixels are crucial in identifying regions with weak textures, assessed by their relative locations, and the identified regions also serve to correspond to the region in stereo pairs. Point set registration is executed for points within the regions. In the matching stage, a matrix form of constraint terms is employed, aimed at ensuring motion consistency across the aggregation of the superpixel. This is complemented by a non-rigid transform model, which plays a pivotal role in directing the registration process. Our motivation is to avoid ambiguities introduced by feature descriptors during the matching process and to limit the distortion of point set alignment by utilizing region-based and locally consistent alignment strategies. Experimental results show that the method successfully matches significantly more keypoints and has the best uniformity performance compared to the comparison methods. The method achieves a more competitive performance in terms of both visual quality and objective assessment compared to other classical feature matching methods and deep learning-based methods.