Maximum entropy-based semi-supervised learning for automatic detection and recognition of objects using deep ConvNets

摘要

Object detection and localisation is one of the major research areas in computer vision that is growing very rapidly. Currently, there is a plethora of pre-trained models for object detection including YOLO, mask RCNN, RCNN, fast RCNN, multi-box, etc. In this paper, we proposed a new framework for object detection called 'maximum entropy-based semi-supervised learning for automatic detection and recognition of objects'. The main objective of this paper is to recognise objects from a number of visual object classes in a realistic scene simultaneously. The major operations of our proposed approach are preprocessing, localisation, segmentation and object detection. In the preprocessing, three processes, noise reduction, intensity normalisation, and morphology are considered. Then localisation and object segmentation is performed using maximum entropy in which optimal threshold is detected and in the end, object detection is performed using deep ConvNet. The performance of the proposed framework is evaluated using MATLAB-R2018b and it is compared with some previous state of the art techniques in terms of localisation error, detection and segmentation accuracy along with computation time.