Mask Refinement for Vehicle Camera Soiling via Linear Modeling of Border Intensity Transitions

Accurate annotation of vehicle camera soiling is critical for developing robust vision-based blindness detection systems in automotive applications. This paper presents a method for refining existing soiling annotations using a combination of Gaussian filtering and adaptive thresholding techniques. Starting with baseline segmentation annotations that exhibit a correlation of 0.61 with grayscale intensity values, the refinement process focuses on enhancing inter-class separability, particularly in border regions between annotated classes. A parameter optimization framework is introduced, wherein Gaussian filter parameters and thresholding thresholds are tuned to maximize the mean intensity difference between classes within the border area. This optimization enhances the distinction between soiled and clean regions, ensuring more precise annotations. The proposed methodology demonstrates improvements in annotation quality, validated by a neural network-based segmentation study. This refinement pipeline provides a robust solution for addressing inaccuracies in vehicle camera soiling annotations, enabling better dataset preparation for training machine learning models.