CNN Image Stitching for Small-Overlap X-Ray Imaging

The patent applies preliminary action by training a convolutional neural network in advance to recognize and predict displacement patterns in overlapping image areas. The pre-trained CNN model can accurately determine the correct displacement between adjacent X-ray images even with minimal overlap, eliminating the need for large overlapping areas that would otherwise be required to ensure stitching reliability. This preliminary training enables the system to achieve reliable stitching results with reduced patient dose.

The patent applies parameter changes by transforming the traditional approach of relying on large overlapping areas into a parameter-based solution using a trained CNN model. The system changes from using spatial overlap extent as the key parameter for reliability to using learned displacement parameters from the neural network. This parameter transformation allows accurate displacement determination even with minimal overlapping areas, maintaining stitching reliability while reducing patient dose.

The patent applies mechanics substitution by replacing the traditional mechanical/optical similarity measure approach with a neural network-based system. Instead of using mathematical similarity computations that are prone to local maxima and minima, the system uses a CNN model that has learned displacement patterns from training data. This substitution maintains computational efficiency while dramatically improving displacement estimation accuracy by avoiding the local extrema problems inherent in similarity-based methods.