Non-Uniform Illumination Image Thresholding Using MLP

One well-known and popular method for classifying objects and backgrounds in real time is thresholding. Because of its simplicity, it is especially well-suited for real-time applications. However, because background and object pixels have overlapping gray-level distributions, thresholding loses effectiveness when used on images taken in uneven lighting. In images with non-uniform lighting, global thresholding techniques fail to binarize because of an absence of visible bimodal histograms. The choice of the initial window size or sub-image size, which differs for each image, affects how well window-based or image partition-based local thresholding performs. Furthermore, the bimodal testing criterion function is necessary for the partitioning process to be effective. Choosing the right criterion function for bimodality and beginning window size for image partitioning are important issues in adaptive thresholding techniques. The suggested approach takes into account machine learning (ML) techniques to increase the accuracy of uneven light image thresholding. The objective is to use global thresholding to binarize an image and predict the non-uniform illumination surface. The suggested MLP-based approach outperforms the local and surface-based thresholding methods in terms of average performance metric.