Proximity Modulated Thresholding for Hessian Matrix Feature Detection

Abstract

Greg Passmore

Our lab processes large volumes of multispectral drone data. We use feature identification for image alignment, object recognition, and scene reconstruction. Traditional methods using the Hessian matrix detect features like corners or blobs. This method is commonly used for color images, and the issues with changes in lighting and exposure are well known. However, for our multispectral data, it proved especially problematic. Fixed thresholds worsened these issues, causing inefficiencies and inaccuracies in feature matching and image alignment.

This paper presents a dynamic thresholding approach that adjusts the feature detection threshold based on the disparity between the current and pre-defined feature count. It starts with an initial detection phase using a standard threshold to establish a baseline. The threshold is then adjusted incrementally until the feature count converges towards the target. This iterative refinement improves responsiveness and efficiency by considering the proximity between the current and desired counts.

Experimental results demonstrate that adaptive thresholding reduces computational costs on the order of one magnitude and can increase the granularity in feature detection processes, making it useful for complex image processing tasks. This approach is particularly beneficial in environments with significant variability in multispectral environments, or where image quality and lighting conditions present challenges.

PDF