REAL-TIME AUTOMATED DEFECT DETECTION IN INDUSTRIAL MANUFACTURING USING DEEP LEARNING AND TEXTURE ANALYSIS

Abstract

Visual inspection is a critical quality gate in industrial manufacturing, yet manual inspection is slow, tiring, and inconsistent, and classical machine vision struggles with variable appearance. This paper presents a real-time automated defect detection framework that combines deep learning with texture analysis. A controlled imaging station with a ring light, a camera, and a tray that holds many items at once provides repeatable images. Three processing modules follow: image pre-processing by grayscale conversion, Sobel edge detection, and binarization, which emphasizes texture and edge structure; target object recognition with a YOLO detector that localizes the objects of interest; and defect detection by measuring the percentage by which an object's extent exceeds a normal baseline. In parallel, a four-block convolutional neural network classifies individual items as good or defective from their appearance, and texture descriptors based on gray-level co-occurrence, local binary patterns, and Gabor responses are fused with its output. On a balanced test set of 144 tablets, the convolutional network alone classified 139 correctly, an accuracy of 96.5% with a precision of 100% and a recall of 93.1%, with no good tablet rejected and five defective tablets missed. Fusing texture features reduced misses from five to two and raised accuracy to 97.9%. Pre-processing raised the object recognition rate under glare from 85.7% to 96.9%, and the complete pipeline processed an image in about 20 milliseconds on a graphics accelerator and about 68 milliseconds on an edge device. The results indicate that combining learned and handcrafted texture representations gives accurate, fast, and robust inspection suitable for production lines.