Abstract:

Objective The industrial sewing robot based on contour extraction detects fabric edge contours with visual aids and works out the robot's movement trajectory based on the fabric edge contour information to achieve the sewing of the fabric in conjunction with the sewing machine. However, the large number of raw edges of the fabric after cutting, the print pattern of the fabric and the background of the fabric image acquisition all affect the accuracy of the fabric edge contour extraction, and the extraction time directly affects the sewing efficiency.

Method The conventional VGG-UNet model was optimized by convolutional splitting and fusion loss functions to improve the inference speed and segmentation accuracy of the model. The optimal fabric detection model was then constructed and trained using the optimized VGG-UNet to segment quickly and accurately the printed fabric and the desktop background, and the fabric burrs were removed using adaptive open operations before the Canny operator was used for edge detection to obtain accurate fabric edge contours.

Results The optimized VGG-Unet optimal training results were 0.79%, 0.79%, 1.6%, 0.79% higher than those of the VGG-UNet model in each index, and the inference speed was reduced by 10.368 ms, and the number of total parameters of the optimized VGG-UNet model was greatly reduced. The optimal fabric detection model that was trained and constructed showed obvious advantages in terms of memory resource consumption and detection efficiency. The superimposed image showed that the contour extraction accuracy was not affected even though the printed fabric was similar in color to the desktop background and the lighting was not uniform. In addition, the contour lines were hand-drawn on the original image, and the fabric edge contour lines extracted in this paper were computed by OpenCV to find out the overlap of the two contour lines, and the accuracy of the contour line extraction was more than 99%. The complete algorithm was obtained by pytorch programming on a computer with Windows 11 operating system, GPU using NVIDIAGerforce GTX 1650 and 16 G memory, and it took only 0.216 s to extract the edge contour of the fabric in a fabric image, while the proposed conventional contour extraction method took 2.852 s. The edge contour of the printed fabric was the worst when the fabric color is close to the background color of the desktop and when the reflection of the desktop was severe. In addition, the conventional contour extraction method does not consider the burr problem generated by the fabric cutting, so the conventional contour extraction algorithm not only has a long extraction time but also cannot remove the noise and the burr efficiently, making it difficult to accurately extract the edge contour of the printed fabric.

Conclusion This paper proposes the use of deep learning combined with conventional contour detection algorithms to extract fabric edge contours for the first time. It solves the problem that traditional fabric contour extraction methods are affected by fabric color, print pattern, fabric texture and desktop background, and has excellent performance in extracting edge contours of fabrics with complex prints. In this paper, we consider the large number of raw edges generated by the fabric edges after cutting the fabric. This method can effectively remove the raw edges and extract the fabric edge contours quickly and accurately, and the extraction process is not affected by the print pattern, table background, color, fabric texture and light source, the method is good in generality and the extraction results fit the fabric edge contours highly.

Key words: printed fabric, edge contouring extraction, machine vision, deep learning, VGG-UNet model, sewing quality