Paper Title
KNITINSPECT: REAL-TIME KNITTING ANOMALY DETECTION POWERED BY COMPUTER VISION

Abstract
Circular knitting machines operating at approx. 30 RPM generate fabric at speeds that outpace human visual perception. A single broken needle can damage 5 to 10 meters of fabric in under sixty seconds, creating a need for automated, low-latency intervention. In this Phase I study, we engineered KnitInspect, a computer vision prototype designed to detect texture anomalies without relying on labeled defect data. We utilized a sliding-window architecture that processes fabric texture via a pre-trained ResNet-18 feature extractor. To avoid the logistical bottleneck of collecting thousands of ”defect” images for supervised training, we implemented an unsupervised Isolation Forest. This approach models the statistical distribution of ”valid” fabric and flags outliers using a dynamic 3-sigma threshold. Testing was conducted on a live industrial setup using a Full HD (1920×1080) GigE camera at 45 FPS. Preliminary validation on a manually collected dataset yielded a detection accuracy of roughly 60% for structural defects. However, latency analysis revealed a critical hardware limitation: processing time on the edge GPU (RTX 3050 Ti) averages 55ms per frame. This exceeds the 22ms actuation window required to halt the machine, establishing the primary optimization metric for Phase II. Keywords - Knitting, Anomaly Detection, Computer Vision, Resnet-18, Isolation Forest, Smart Textiles, Quality Control