Mathematical Modeling and Algorithmic Optimization for Intelligent Signal and Image Processing Applications

Abstract

The image segmentation is a very important process in computer vision and signal processing, but LiDAR-acquired height images are rather challenging because of noise, irregular signal patterns, and different object-background relationships between objects and the background. The traditional threshold-based segmentation techniques tend to be ineffective in such conditions. This study developed and evaluated an optimized segmentation framework for LiDAR-derived height images using mathematically grounded and computationally efficient image-processing techniques. A dataset of 120 annotated LiDAR height images was analyzed under three noise conditions: clean, medium, and heavy. The baseline approach applied Gaussian smoothing followed by Otsu thresholding, while the optimized pipeline incorporated image normalization, Gaussian filtering, adaptive threshold polarity selection, morphological refinement, and connected-component filtering. Segmentation performance was evaluated using the Dice similarity coefficient and Intersection over Union (IoU). The baseline method showed extremely limited performance, with mean Dice and IoU scores of 0.0173 and 0.0097, respectively. In contrast, the optimized segmentation framework achieved mean Dice and IoU values of 0.7903 and 0.6673. Strong segmentation accuracy was observed for clean and medium-noise images, while performance remained acceptable under heavy-noise conditions. These findings demonstrate that algorithmic optimization can substantially improve segmentation accuracy for LiDAR-derived height images while maintaining interpretability and computational efficiency, providing a practical alternative to more complex learning-based approaches for specialized datasets.