An Entropy-Based Feature Selection Algorithm for Curve Detections and Parameterization in Grayscale Images
Tamás Reiter, Michael McCann, Nathan Henry · 2024
This paper presents a local-global entropy-based algorithm for image edge band localizations. The proposed method computes a series of locally filtered entropy images at variable window sizes and their corresponding global entropy values. Furthermore, the optimal entropy-filtered image is identified by selecting the one with the lowest corresponding global entropy value hypothetically ensuring the lowest potential disorder or variance within the image. Evaluations are conducted comparing the proposed method with the most popular edge detection techniques using a limited set of industrial semiconductor images, demonstrating competitive results. The proposed algorithm is interesting, input parameter free and integrable with other techniques. However, the computational costs may increase significantly depending on the number of local entropy window sizes that need to be processed. Our algorithm can also be extended for future research to incorporate adaptive local window sizes within the image. This algorithm is not designed to compete with state-of-the-art traditional and deep learning-based edge detectors. Instead, it aims to serve as a reliable, manual-tuning-free feature selection method for grayscale images, particularly in scenarios where intricate intensity values contain rich informational content for measurement analyses.