A Comprehensive Review on Various Image Segmentation Methods with Applications

The basis of numerous computer vision applications is image segmentation, which entails dividing an image into coherent areas. In this study, different methods that are explored for image segmentation are introduced, and their strengths and weaknesses are discussed. We start by emphasizing the importance of correct segmentation for the object detection, recognition, and feature extraction tasks. Tensor representation of various types of image data, then we investigate into the popular field of segmentation techniques which can be mainly grouped as thresholding, region-based, edge-based and the deep learning methods. Principles, advantages, and weaknesses for each category are detailed. Thresholding techniques are very simple and computationally very quick and are noise and uneven illumination sensitive. Region-based approaches like region growing and merging are good in extracting connected regions, but cannot deal well with the overlapping objects. The edge-based methods using the intensity discontinuity are vulnerable to edge detection errors and do not include spatial information. Deep Learning methods based on convolutional neural networks showed impressive accuracy but still depend on vast amount of training data and on high computational power. Our critical analysis of these divergent categories helps to illuminate possible trade-offs related with each approach. Our study stresses the necessity to adapt the segmentation method to the particular application and image properties. Finally, we discuss the trends and perspectives of such developments, which will inspire the performance and generalization of image segmentation.