Over the past 3 years, a dedicated AI development team at Shelton Vision has aimed to elevate the detection process and the accuracy of naming and grading subtle defects in textiles, in real time within production environments.
Big Data ‘off-the-shelf’ systems involve reading many thousands of single images each second and simply take too long to accumulate sufficient data for what’s required in this specific case. In many sectors of the textile industry, the product range changes several times within a year, and it is not uncommon to have to inspect hundreds, if not thousands of different styles in a year based on precise settings. In terms of defect types, there may typically be over 100 that need to be accurately detected, classified and graded in real time.
There is also the need to ‘filter out’ the random occurrence of ‘non defects’, such as loose threads, lint and dust on the surface, and it is clear that a bespoke system is required. The development team has therefore established metadata for identifying defect properties, enabling the successful identification of faults from a much smaller number of images.
The system employs a unique combination of machine learning for automated style training and novel algorithms for defect detection, to provide high quality images for the AI real time defect classification and grading software. Due to the inherent variation in fabric features – raw materials, construction, texture, color and finishes, as well as the differing product quality standards in value chains and the regional variations in what defects are called – the AI engine uses models built for each individual company or group of companies, or product value chain.
The AI models from Shelton Vision Ltd., Leicester/UK, are constructed so that the user operatives can populate them with their own data produced by the vision system or by obtaining defect images from another imaging source (e.g. a mobile phone camera). The occurrence of defects is sporadic and many defect types occur infrequently. The AI engine can be quickly set up and able to operate accurately with limited data sets of typically between 30 and 50 good quality images per defect type.
A further feature is a tool enabling the user to periodically ‘clean up’ the AI data during the set-up phase. This is used to resolve conflicting data and to correct mis-named images.