Performance improvement on object detection using deep learning with image enhancement techniques

Abstract

Object Detection intends to localize and segment the most apparent objects or regions in an image. It is widely used in visual applications like object re-targeting, object classification, image synthesis, object tracking, image retrieval, etc. The main complications encountered by an object detection algorithm are non-uniform illumination, various postures, occlusion etc., which cause false object detection; Because of complex background, the accuracy of tradition object detection algorithms will drop sharply. However, there is an increasing demand of deep learning approaches for object detection. In deep learning, when training to detect the object, the images must have good quality. Images from the datasets are taken in different illumination conditions. Low illumination properties leads to loss of information, which in turn makes it hard to detect the objects. Distinct from existing detection methods, which conduct object detection directly on original degraded images, the thesis eliminates the effect of low illumination images or degraded images by an explicit enhancement of the image. There is a great demand for image enhancement to solve different applications in various fields. The image enhancement mechanism consists of various techniques that are used to enhance the appearance of an image by eliminating blur and noise in the image. These techniques enhances the geometric features of objects like edges and also the classification and detection performance of the machine learning models. The various image enhancement models discussed in the literature are histogram equalization, gamma correction, contrast limited adaptive histogram equalization, etc. newline