Image fidelity assessment to support in-process deep learning based object detection

The effectiveness of Deep Learning (DL) models for object detection depends heavily on the fidelity of training images, especially in industrial defect detection applications. Image fidelity, defined as the accurate representation of an object in a digital image relative to human perception, is particularly critical in detecting amorphous defects, such as wrinkles on automotive seats. A comprehensive literature review highlights a significant gap in assessing and optimising image fidelity for defect detection applications, leading to inconsistencies in DL model performance. This research addresses this gap by developing a Physical Product Labelling System (PPLS), a novel method for quantitatively assessing image fidelity to enhance machine vision-based defect detection. PPLS consists of three key components: (1) a Physical Product Labelling Method (PPLM), which introduces a structured workflow for inferring I §mage fidelity; (2) a Physical Product Labelling Tool (PPLT), designed to allow Subject Matter Experts (SMEs) to mark Objects of Interest (OOIs) on a physical product and transform these into 2D image labels; and (3) a validation framework to assess the repeatability and reproducibility of PPLT in real-world industrial applications. The study applies PPLS to detecting wrinkles on leather automotive seats (the OOI), a challenging use case due to the amorphous nature of wrinkles and their sensitivity to imaging conditions. The results demonstrate that lighting conditions suitable for a human inspector and standard machine vision lighting do not provide optimal imaging conditions for the OOIs, underscoring the necessity of an image fidelity assessment framework. The research confirms that quantifying differences between human and digital perception of defects enables the optimisation of imaging parameters. The findings contribute a systematic approach to image fidelity assessment, facilitating the integration of high-fidelity images into DL model training for defect detection. The research also highlights broader implications for machine vision applications in industrial quality control, paving the way for further studies on adaptive imaging systems, real-time projection methods, and industry-wide application of image fidelity evaluation techniques.