Semantic Segmentation of image Using Deep Learning:Review

Semantic segmentation is considered as one of the most important and challenging problem in the field of computer vision which aims at assigning a class label to each pixel in an image which leads to sophisticated scene understanding. This task has been extensively used in various application areas including the self-driving cars, medical diagnosis, and environmental monitoring. Semantic segmentation has come a long way since its early detection algorithms based on features extractions to state of the art deep learning methods.This paper aims at presenting the evolution of the semantic segmentation, and specifically, how the deep learning has changed the field. The conventional approaches consisting of edge detection and histogram analysis offered a basic level of understanding but were constrained by the use of hand crafted features. Deep learning, however, is capable of learning features and has produced very promising results across numerous tasks. Some important architectures that have set the benchmark in the field include Fully Convolutional Networks (FCNs), U-Net, and DeepLab which have used convolutional layers, encoder-decoder architecture, and atrous convolutions for improving the accuracy of the segmentation. The article also reviews some of the publicly available datasets which include Cityscapes, PASCAL VOC and ISIC 2017 which are widely used to assess the performance of the segmentation models. These datasets differ in their complexity, resolution, and the application domain that they cover which makes the problems that they present to researchers diverse. Also, we compare the traditional and deep learning based feature extraction methods and present the characteristics of each method, their advantages, and disadvantages, and areas of application. This survey aims at assisting researchers and practitioners by presenting the current best practice in the form of state-of-the-art methodologies, discussing the potential of application of such methodologies in the real world, and identifying the directions for further research.Therefore, despite the advancement of deep learning in the area of semantic segmentation, there are still numerous issues which need to be addressed in the future, including efficiency, scalability, and domain specific issues. This all-encompassing review paper is hoped to be beneficial to those wishing to gain more knowledge on the current trends as well as find a way to contribute to the field of semantic segmentation in the future.