Deep-feature-matching-based B-spline Scale Correction Method for Nonlinear Distorted Images

Images collected by linear scan cameras are nonlinearly stretched and compressed due to the speed regulation of high-speed trains. This condition changes the shape of objects and thus considerably increases the false and missed alarm rates for the subsequent troubleshooting. Therefore, a four-module image morphology restoration methodology is proposed, in this article, for high-speed running trains. First, deep-feature-based motion smoothness feature matching method, namely DeepGMS, is proposed which can yield out rich features even in low texture, high-light and similar feature confusion situations. Second, a dense feature sparsification algorithm via B-spline fitting is put forward to acquire the best feature points (BFPs) represented various scales. Third, a trajectory planning B-spline interpolation is raised to connect each BFP in series to form a continuous and smooth curve and thus simulating the nonlinear scale of morphology caused by speed changes. At last, the remapping module maps the interpolation parameters above to all pixels in the entire image to achieve correction. Experimental results on the 6 captured image sequences of high-speed trains in operation show that the proposed algorithm can restore the original shape of high-speed trains in real-time (1 second per 2048×2048 image) and outperforms the existing methods in terms of speed and precision.