Image stitching method for CMOS grayscale cameras in industrial applications

Abstract

To address the limited field of view (FOV) of CMOS grayscale cameras, complex lighting conditions, and the scarcity of image features in industrial applications, a novel image stitching method is proposed for CMOS grayscale cameras operating under varying lighting conditions. This method broadens the camera’s FOV while preserving the interpretability of image features, thereby enhancing the robustness and generalizability of image stitching across diverse lighting environments and feature-sparse settings. In the feature extraction phase, a hybrid deep feature extraction network is designed. By employing a deep learning-based approach, the network ensures the extraction of a substantial quantity of features. Building on this foundation, a method for line feature selection and reconstruction is developed to refine feature-matching accuracy, which increases the number of matching lines in extreme lighting and feature-scarce situations, and enriches the image features for subsequent stitching processes. In the subsequent image transformation phase, planar feature constraints are introduced; matching feature points and lines are used to generate planar features, addressing alterations in the collective shape of planes that are common in industrial image stitching. The paper concludes by presenting quantitative evaluation metrics for planar feature-based stitching. Experimental results validate the effectiveness and feasibility of the proposed method for image stitching of CMOS grayscale cameras under varied lighting conditions and in feature-deficient industrial settings, offering a viable solution to the challenges posed by the limited imaging FOV in industrial applications.