Title: Artificial Intelligence Across Disciplines: Research, Innovation, and Intelligent Solutions
Editors: Dr. Subita Bhagat, Dr. A. Balamurugan, Dr. P. Krishna Kumar, and Mrs. S. Nandhini Devi
ISBN: 978-81-69857-83-3
Chapter: 3
DOI: https://doi.org/10.59646/815/03
Author: Vrushali Kailas Jadhav
Abstract
Modern industrial automation requires automated visual inspection capable of detecting microscopic surface defects across high-speed manufacturing lines. Traditional computer vision pipelines rely heavily on hand-engineered morphological filters and thresholding operators, which fail when confronted with specular metallic reflections, texture variations, and severe class imbalance. This chapter proposes a deep visual analytics framework combining a Swin-Transformer-enhanced Feature Pyramid Network with an anchor-free detection head and an orthogonal boundary-refinement branch. Using precision additive-manufactured aerospace alloy inspection as a real-world case study, the architecture captures both multi-scale contextual dependencies and high-frequency edge discontinuities. We formulate a composite objective combining balanced focal loss, generalized intersection-over-union loss, and active-contour boundary regularization. Empirical evaluation demonstrates a mean Average Precision of 88.6% at 50% intersection-over-union, an inference latency of 22.4 milliseconds per multi-megapixel frame, and a 42.7% reduction in false-positive rejections compared to convolutional baselines, confirming its viability for high-throughput industrial defect screening.
Keywords: Computer Vision, Deep Learning, Surface Defect Detection, Vision Transformers, Industrial Quality Control, Feature Pyramid Networks.