Deep-feature-matching-based Piecewise Polynomial Transformation for Multi-domain Nonlinear Scale Distortion Image Correction Method

Images collected by linear scan cameras are stretched and compressed due to the speed regulation of targets. This condition changes the target morphology and considerably affects the accuracy of subsequent image tasks such as target detection and tracking. Affine transformation matrix is the priority in fields of image alignment, registration and correction, to solve the mapping relationship between pixels. However, in practical engineering applications such as aerospace circular scanning camera imaging, through-type scanning high-speed train body fault detection, and industrial circuit board assembly line fault detection, the scanning rate of linear scan camera cannot be adaptively adjusted in real time according to the target speed, resulting in deformations of various scales in the sequence image. A unified affine matrix solution for the entire image cannot solve the problem of target scales with regional differences. Therefore, a deep-feature-matching-based piecewise polynomial transformation for multi-domain nonlinear scale distortion image correction method is innovatively proposed, to achieve one-to-one correspondence between deformed images and standard images at the pixel level. First, a feature matching algorithm based on deep feature extraction, driven by big dataset, is integrated with graph convolution and optimal matching to achieve high-precision matching of two images, which performs well in illumination, weak texture, and similar features confusion. Second, an image block algorithm based on sparse feature clustering is adopted, which applies k-means clustering to sparsely reduce the dimensionality of dense feature points to maintain a good balance between speed and accuracy. Third, a novel image correction method based on piecewise polynomial trajectory planning is proposed, which converts the multi-domain nonlinear scale distortion image correction into robot multi-segment path planning. This strategy optimizes the coordinate standardization formula of each scale-changing image block, and guarantees the continuity of pixels between adjacent image blocks, thereby achieving high-precision image mapping relationship solution. At last, a remapping method of sparse feature to dense pixel is proposed and transforms the original high-resolution deformed image to the corrected one via the calculated mapping formula. Experimental results on 5 long sequences has demonstrated the effectiveness of our methodology with an maximum 1.25% improvement, which can be expanded to remote sensing image panoramic stitching, calibration and registration.