Image Pattern Analysis using Machine Learning

Abstract

The advancement in machine learning techniques and high-speed computing systems newlineboosted the image data analysis and interpretation. Enhanced deep learning methods newlineusing multi-level training such as transfer learning helps to learn better the class newlineproperties. This has enriched the domain of image analysis and insight information newlineextraction. The deep learning technique is used for automatic image feature extraction, newlineimage classification, and image segmentation. The use of high-quality imaging newlinetechniques in the health care system and popularity of image and video data in social newlinemedia has made image analysis one of the most demanding research areas in computer newlinevision. newlineThe objective of this thesis is to develop an enhanced framework (a) for medical image newlineanalysis, (b) to build an image retrieval system using better semantic features of the newlineimage and feature fusion, (c) to design enriched scanned document information newlineextraction systems. The models are developed using deep learning and feature fusion newlinetechniques. Five different datasets are used to test the efficacy of the different models. newlineThe performances of the models are compared with the existing state-of-art using newlinestandard performance measuring metrics. newlineThe medical image analysis we have considered two types of image dataset i.e. the newlinefundus image database for the diabetic classification, and the radiological image newlineclassification for brain tumor and COVID-19 detection. For fundus image analysis the newlinedeep learning technique particularly the UNet is used for the image segmentation. Two newlinedifferent UNets are designed and trained for the segmentation of the blood vessels and newlineoptic-disc segmentation from the fundus image to deal with their complexity in the newlinestructural and visual properties. After the due removal of the blood vessels and the newlineoptic-disc, the deep learning features are extracted and classified by a multi-class SVM newlineclassifier. The classification result is compared with the current-state-art. To analyze newlinethe radiological image a saliency-region enhanced CBMIR sy