Optimization of image mosaic in weak texture scene by region localization and geometric enhancement strategy

INTRODUCTION: In industrial production and equipment manufacturing, there's a strong practical need to create wide-angle panoramic images through image stitching for comprehensive visual monitoring, but actually deploying it still comes with challenges. However, traditional SIFT-based image registration methods often struggle in low-texture scenes because the feature points are sparse and unevenly distributed, which can easily cause mismatches in non-overlapping areas and lead to registration errors or failures. OBJECTIVES: This article introduces a new image stitching framework based on SIFT, aimed at improving the quality and accuracy of image stitching in low-texture scenes. METHODS: We first use a template matching method based on normalized cross-correlation to locate overlapping areas between images, which helps effectively reduce the interference from non-overlapping features and improves the reliability of feature extraction and matching. When the overlapping area mostly consists of weak texture regions, we add a feature matching simulation that uses successfully matched key points to extend extra geometric constraints in the overlap to better link the images. RESULTS: The experimental results show that the proposed method performs well when stitching images with large areas of weak textures, significantly improving both registration accuracy and overall stitching quality. CONCLUSION: The proposed method has been initially tested in several challenging scenarios like low light, strong backlighting, and infrared, showing its potential for robust image stitching in low-texture environments.