Paper Key : IRJ************066
Author: Ronav Jaiswal,Nilesh Jaiswal
Date Published: 24 Oct 2023
Abstract
Machine learning based object detection and classification in digital images usually suffers from accuracy issues as the model often gets trained on background pixels and object textures that vary significantly between training and classification stages. In such cases the model identifies objects incorrectly and results in higher than acceptable false-positive and false-negative rates. In this paper the authors propose a novel stylize-cluster-classify approach to improve feature extraction and classification accuracy during object detection in images. The approach has been tested with a custom "Blind Objects" image dataset that consists of everyday objects encountered by people. This research has applications not just for visually impaired persons to identify objects accurately but also in related domains such as medical image prediction and novel body detection in astronomy images. In this approach the authors propose two additional stages in the machine learning workflow: (1) stylization and (2) clustering. Stylization transfers specific rendering styles to the dataset images in order to reduce histogram spread and pixel noise, in addition to improved detection of image features. The clustering stage improves classification accuracy by introducing an additional dimension of cluster identifier to the image feature vector. This additional dimension has a small negative effect on algorithm performance but results in a large improvement in accuracy.
DOI LINK : 10.56726/IRJMETS45537 https://www.doi.org/10.56726/IRJMETS45537