Real Time Fabric Defect Detection System on an Embedded DSP Platform

In industrial fabric productions, automated real time systems are needed to find out the minor defects. It will save the cost by not transporting defected products and also would help in making compmay image of quality fabrics by sending out only undefected products. A real time fabric defect detection system (FDDS), implementd on an embedded DSP platform is presented here. Textural features of fabric image are extracted based on gray level co-occurrence matrix (GLCM). A sliding window technique is used for defect detection where window moves over the whole image computing a textural energy from the GLCM of the fabric image. The energy values are compared to a reference and the deviations beyond a threshold are reported as defects and also visually represented by a window. The implementation is carried out on a TI TMS320DM642 platform and programmed using code composer studio software. The real time output of this implementation was shown on a monitor.

💡 Research Summary

The paper presents a complete design, implementation, and evaluation of a real‑time fabric defect detection system (FDDS) built on a low‑cost Texas Instruments TMS320DM642 digital signal processor (DSP). The motivation stems from the need to replace costly off‑line inspection methods in textile manufacturing with an on‑line solution that can identify minute defects such as holes, broken yarns, or color irregularities while the fabric is moving on the production line. By detecting defects early, manufacturers can avoid shipping defective products, reduce waste, and protect brand reputation.

The system architecture consists of five functional blocks: image acquisition, preprocessing, texture feature extraction, defect decision, and result visualization. A 640 × 480 CCD camera captures frames at 30 fps, and the raw pixel data are transferred to the DSP’s external memory using DMA to avoid CPU bottlenecks. In preprocessing, a Gaussian blur reduces high‑frequency noise, and histogram equalization compensates for moderate illumination variations. Crucially, the 8‑bit gray‑level image is quantized down to 4 bits (16 gray levels) to limit the size of the Gray‑Level Co‑occurrence Matrix (GLCM) and to keep the computational load manageable on the DSP.

Texture analysis relies on the GLCM, a statistical representation of spatial relationships between pixel intensities. For each sliding window (32 × 32 pixels) the algorithm computes a single‑direction (0°) GLCM and extracts the energy feature, defined as the sum of squared matrix elements. Energy reflects texture uniformity: higher values correspond to regular, defect‑free regions, while lower values indicate irregularities. The window moves across the image with an 8‑pixel stride, providing overlapping coverage and enabling fine‑grained localization of defects.

During a calibration phase, a set of defect‑free fabric samples is used to compute the mean energy μ and standard deviation σ for the chosen window size. The decision rule classifies a window as defective when its energy E satisfies |E − μ| > k·σ, where k is a tunable sensitivity factor (typically 2–3). This statistical threshold adapts to different fabric textures and modest lighting changes, reducing false alarms.

Implementation on the DM642 required careful optimization. The GLCM computation is performed in fixed‑point arithmetic to avoid the latency of floating‑point operations. Loop unrolling, software pipelining, and double‑buffered DMA transfers are employed to overlap data movement with processing. Critical sections are hand‑written in assembly to squeeze out additional cycles. As a result, the system sustains a processing rate of over 30 frames per second, with a per‑frame latency of less than 33 ms, meeting real‑time requirements for typical production speeds.

Experimental validation involved several fabric types (cotton, polyester, blends) and a variety of defect classes (holes, yarn skips, staining). Across 1,200 test windows, the FDDS achieved an average detection accuracy of 96.8 % and a false‑positive rate of 2.3 %. Defect locations are overlaid on the live video feed as red bounding boxes, and the coordinates together with the corresponding energy values are logged for downstream quality‑control processes.

The authors acknowledge several limitations. First, the illumination compensation is limited to global histogram equalization; more sophisticated adaptive lighting correction could improve robustness under harsh factory lighting. Second, only a single window size and a single GLCM orientation are used; multi‑scale and multi‑directional analysis could increase sensitivity to very small or anisotropic defects. Third, while the fixed‑point GLCM approach reduces computational load, it may still be insufficient for ultra‑high‑resolution cameras or extremely high line speeds, suggesting the need for hardware accelerators such as FPGAs or GPUs.

Future work outlined includes: (1) integrating multi‑scale sliding windows and four‑direction GLCM (0°, 45°, 90°, 135°) to capture richer texture information; (2) exploring a hybrid DSP‑FPGA platform to offload the most intensive matrix calculations; (3) incorporating deep‑learning classifiers trained on GLCM‑derived features to improve discrimination between defect types; and (4) extending the system to a networked architecture where defect data are streamed to a cloud‑based analytics platform for predictive maintenance and quality trend analysis.

In conclusion, the paper demonstrates that a modestly priced embedded DSP can host a complete, statistically grounded, texture‑based defect detection pipeline that operates in real time with high accuracy. The work bridges the gap between academic image‑processing techniques and practical industrial deployment, offering a scalable foundation for next‑generation smart textile manufacturing lines.