AI System for Anomaly Detection and Quality Control of Fabrics

Below is a list of the hardware equipment used:

• Camera: GoPro HERO13 Black
• Lens: Macro lens mod
• Remote control: Remote camera controller

The camera was initially positioned in the center, far from the fabric, and then about 15 cm away from it, to frame an illuminated portion of material. For this purpose, the illuminator (an existing lamp within the system) was oriented in alignment with the camera so as to minimize shadows and undesired effects. The installation was carried out near the core of the machinery, thereby reducing visual artifacts due to fabric transparency, which would otherwise reveal external elements, such as operators, objects, and the surrounding environment.
The use of the macro lens allowed close-up shots (range 11–75 cm), ensuring adequate resolution for defect inspection. The camera was fixed to a structural part of the machine anchored to the ground, thus not subject to the rotation of the tubular machine. In this way, the fabric itself moves around the field of view, allowing the use of a single camera to capture the entire surface in real time.
The use of the remote control made it possible to take photos independently, while through the GoPro Quick app it was possible to configure the camera, check the framing, and download images in real time, even after the installation.
For automatic acquisition, the time-lapse mode at regular 5-second intervals was used, which allowed continuous sampling of production.

The automatic dataset generation process, essential for training the described model, follows a pipeline with the following main steps.

• Extraction of compliant patches: multiple anomaly-free patches are derived from each acquired image of compliant fabric.
• Introduction of synthetic defects: artificial defects are applied to some patches, generated by randomizing parameters such as shape, color, and texture, to simulate realistic anomalies, which are then realistically merged into the patches.
• Automatic annotation creation: since the defects are inserted in a controlled manner, the corresponding segmentation mask (target) is automatically produced, without the need for manual labeling.

This strategy made it possible to obtain large, high-quality datasets with significantly reduced time and costs compared to a traditional annotation process, thus constituting a key element for the scalability of the proposed solution.

System Evaluation and Results

To evaluate the system's performance, we built a test dataset comprising 200 image crops (256x256 pixels). These crops were extracted exclusively from images not used during the model's training phase. The test set is composed of:

• 100 defective samples (BAD Class): manually selected crops, each containing at least one real visual defect (e.g., tears, stains, irregularities in the fabric pattern, …); see Figure 1.
• 100 normal samples (GOOD Class): visually intact fabric crops, representing standard production quality; see Figure 2.

For each test sample, the model outputs an anomaly map with values ranging from 0 to 1. This map has the same dimensions as the input crop (256x256 pixels), where each pixel's value is proportional to the estimated probability of an anomaly in that specific region. High values indicate areas the model considers anomalous, while low values point to normal areas. To condense this two-dimensional map into a single numerical value, necessary for classifying the sample as defective (BAD) or not (GOOD), we calculate a specific percentile of the map's values. For the tests detailed in this document, we used the 97.5th percentile.

Classifying a sample as BAD or GOOD requires setting a threshold on the anomaly score: samples scoring above this threshold are flagged as defective, while the rest are considered intact. However, the optimal threshold heavily depends on the specific application context and system requirements (e.g., the acceptable trade-off between false negatives and false positives).

Therefore, we chose to evaluate the model's performance using threshold-independent metrics, which describe the system's behavior across all possible thresholds. The selected metrics are AUROC and AUPRC:

• AUROC (Area Under the ROC Curve): The ROC curve plots the True Positive Rate (TPR, the fraction of correctly identified defective samples) against the False Positive Rate (FPR, the fraction of intact samples incorrectly flagged as defective) across various thresholds. The AUROC is the area under this curve, ranging from 0 to 1. A value of 0.5 corresponds to a random classifier (equivalent to a coin toss), whereas 1.0 indicates perfect class separation. Practically, AUROC can be interpreted as the probability that the system assigns a higher score to a randomly chosen defective sample than to a randomly chosen intact one.
• AUPRC (Area Under the Precision-Recall Curve): The Precision-Recall curve illustrates the trade-off between precision (the fraction of samples flagged as defective that actually are) and recall (the fraction of actual defects successfully identified by the system) across different thresholds. The AUPRC is the area under this curve and also assumes values between 0 and 1. This metric is particularly insightful for real-world applications, as it provides a single-number summary of the system's ability to simultaneously minimize both false positives and false negatives.

The system recorded an AUROC of 99.93% and an AUPRC of 99.93%. These results indicate the model's capability to separate the two classes, assigning higher anomaly scores to defective samples while minimizing both false positives and false negatives across potential operational thresholds.

While threshold-independent metrics are our primary focus, we also evaluated the system's performance at a specific operational point using secondary metrics. We used the default threshold of 0.5 (since the anomaly score ranges from 0 to 1). Applying this threshold yields an accuracy of 99.50%.

Figure 3 shows two additional plots that are useful for effectively understanding the model's performance on the test set.

• Figure 3 (a) illustrates the trends of the True Positive Rate (TPR) and True Negative Rate (TNR) as the anomaly threshold varies. Analyzing this curve helps identify the system's optimal operating point based on specific operational requirements.

Figure 3 (b) displays a boxplot of the anomaly score distributions for the two test dataset classes. This visualization intuitively highlights the separability between GOOD and BAD samples. The further apart and less overlapping the two distributions are, the better the system can sharply discriminate between intact and defective fabric.

Figure 4 presents qualitative results for selected samples from the test set. For each displayed sample, the following elements are provided:

• The original input image, marked with a red dot in the top-left corner if the ground truth is BAD, or green if it's GOOD.
• The anomaly map (using the magma colormap) overlaid transparently on the input image.
• The system's anomaly score as a percentage (0%–100%).

Staff Jersey’s average score shows it is already at a moderately advanced stage of its digital transformation process, outperforming above SMEs of the textile sector across Italy and Europe. The company’s current investments in digital technologies cover a wide range of business operations, with some space to increase its level of preparedness (in terms of plans and resources) in order to accommodate more sophisticated solutions. Staff Jersey uses most of the available mainstream technologies for its operations and looks forward to adopting disruptive technologies (i.e. AI), which is the core element of the EDIH service it applied for.

The company’s personnel has an increased level of digital skills, but needs a well planned and executed training programme and IT specialised professionals to support its potential. Data management capacities and data security are advanced but could further benefit from digital technologies that would bring the business intelligence needed in order to outperform competition.