Convolutional neural networks and multi-threshold analysis for contamination detection in the apparel industry

TL;DR

This paper presents a combined multi-threshold analysis and deep learning approach for automatic contamination detection in apparel manufacturing, achieving high accuracy in real-world industrial settings.

Contribution

It introduces a novel two-level processing method combining threshold analysis and deep learning for contamination detection in textile products.

Findings

False negatives below 3%

False positives below 15%

Successful deployment in a production plant

Abstract

Quality control of apparel items is mandatory in modern textile industry, as consumer's awareness and expectations about the highest possible standard is constantly increasing in favor of sustainable and ethical textile products. Such a level of quality is achieved by checking the product throughout its life cycle, from raw materials to boxed stock. Checks may include color shading tests, fasteners fatigue tests, fabric weigh tests, contamination tests, etc. This work deals specifically with the automatic detection of contaminations given by small parts in the finished product such as raw material like little stones and plastic bits or materials from the construction process, like a whole needle or a clip. Identification is performed by a two-level processing of X-ray images of the items: in the first, a multi-threshold analysis recognizes the contaminations by gray level and shape attributes; the second level consists of a deep learning classifier that has been trained to distinguish between true positives and false positives. The automatic detector was successfully deployed in an actual production plant, since the results satisfy the technical specification of the process, namely a number of false negatives smaller than 3% and a number of false positives smaller than 15%.