The Influence Measures of Light Intensity on Machine Learning for Semantic Segmentation

Abstract

For the human eye, the conversion of light intensity through optic nerve is a non-linear conversion. Therefore, the differences of color caused by light intensity will be reduced by this mechanism. However, the conversion of light for the photosensor in camera is linear conversion, which also causes great influence on the image. Semantic segmentation could be known as a pixel-wise classifier. This technique can be implemented by machine learning or deep learning. In deep learning, the difference in light intensity has a relatively low impact because of relatively strong learning ability. For machine learning algorithms, it will have a significant impact because the classification method is based on RGB values. In this study, the light intensity of the training data would be calibrated and then the random forest model trained from the processed datasets would be compared with the model trained from the unprocessed datasets.