关键词: 机织物, 静水压, 抗渗水性, 水珠, 高斯混合模型

Abstract: Hydrostatic pressure resistance of textiles is an important indicator affecting the wet comfort of textiles. In fabric research and testing stage, the hydrostatic pressure method is commonly used to assess the water resistance of textiles. Current standards such as ISO 811:2018, GB/T 4744—2013, and AATCC 127—2017 are applicable to evaluating the water resistance of various fabrics and non-woven materials (such as canvas, geotextiles, and tent fabrics) that have undergone waterproofing treatments. However, these standards still require inspectors to stop the equipment when the third water droplet is observed. Manual judgment has many disadvantages, such as the delay in human-machine operation, the inability to accurately describe the water discharge position, the need for inspectors' presence, and poor reproducibility. Therefore, exploring the automatic ispection of the hydrostatic pressure of woven fabrics is of great significance. Machine vision-based hydrostatic pressure testing can be understood as dynamically tracking transparent, nearly circular water droplet targets on the substrate of fabric. Currently, existing methods for detecting moving targets include optical flow, frame difference, and background subtraction.
There are still some shortcomings in the current image-based detection of dynamic water droplets on fabrics. First, optical flow method has a high computational complexity, which can easily lead to delay and misjudgment in video droplet tracking. Second, the frame difference method is sensitive to light and holes are easy to appear in the segmented motion foreground when water droplets move slowly. There are many limitations in its application. Third, Gaussian mixture model has weak convergence and poor contour detection integrity, and is not robust to external factors such as environmental noise and lighting. Fourth, infrared images have poor detection results in static water pressure testing due to the small temperature difference between water droplets on the fabric surface and the fabric surface caused by prolonged contact. To improve the efficiency of static water pressure testing of woven fabrics and verify the effectiveness of the image analysis method in detecting the static water pressure of fabrics, we adopt a machine vision-based automatic detection method for fabric hydrostatic pressure. By utilizing 3D printing technology, the encapsulation of the acquisition equipment and light source is achieved. Real-time masking, denoising, and segmentation processing of video frames are performed to obtain a stable and effective observation area. By using the  background subtraction method improved by the background updating strategy, and combining with a mixture of Gaussian models, we achieve the real-time recording of the water outlet point position of the fabric and frame number, which can be used to calculate the fabric's resistance to static water pressure. We also develop a dynamic detection system that can monitor fabric hydrostatic water pressure, automatically stop testing, extract keyframe images, record time, and calculate the fabric hydrostatic pressure. The system include four modules of video image acquisition, pre-processing, motion droplet detection and data recording conversion.
To verify the adaptability of the proposed testing method, experiments were conducted and compared with the existing equipment's built-in detection module, conventional background subtraction method, and Gaussian mixture model subtraction method. Compared with existing methods, results show that this improved algorithm performs well in detecting fabrics with solid color or wide stripe, and the errors range from 0.37% to 2.77%. However, for fine stripes and irregular printed fabrics, the error rate is higher, being above 9.27%. This method can effectively detect solid color and some regular patterned fabrics, but its applicability to complex textured fabrics needs to be improved.

Key words: woven fabrics, hydrostatic pressure, water resistance, water droplets, gaussian mixture model