Investigation and Development of Feature Extraction Algorithms for Object Detection and Recognition in Images

Abstract

Computer vision finds its applications in a variety of fields which includes scene newlineunderstanding, computational photography, medical image processing, tele-medicine, satellite image processing, autonomous cars, Advanced Driver Assistance Systems (ADAS) etc. Identifying the best features for object recognition is a significant area of study and research, and is vital to many applications in computer vision. Object detection is a computer vision newlinetechnique that allows users to identify and locate objects in an image or video. With this kind of identification and localization, object detection can be used to understand the scene, count objects in a scene and determine and track their precise locations, all while accurately labelling them. newlineIn a broader perspective, object detection can be accomplished by using two approaches: newlinemachine learning-based and deep learning-based. In more traditional machine learning-based approaches, computer vision techniques extract various features of an image, such as the texture, shape, colour histogram or edges, to identify groups of pixels that may belong to an object. These features are then fed into models that predicts the location of the object along with its label. On the other hand, deep learning-based approaches employ Convolutional Neural Networks (CNN) to perform end-to-end, unsupervised object detection, in which feature extraction is not performed trivially. In this thesis, both these approaches have been deployed newlinefor object detection and recognition. newlineExtraction, identification and selection of best features for detection and extraction of specific objects of interest (face, traffic sign) from the given scene represented as images is the key behind this research. In order to accomplish this, primarily a set of algorithms are required to appropriately handle real world scenarios with their goal to detect, recognize and classify newlinephysical objects defined as interest by the users. A few algorithms have been proposed in this thesis to identify the best features for a given use..