关键词: 缝纫线迹, 质量检测, YOLOv4-Tiny, 卷积注意力机制, 快速空间金字塔池化, 服装质量

Abstract:

Objective It is reported that manual quality detection of sewing stitch is inefficient and that the existing algorithms are difficult to detect the sewing stitch and the detection is easily interfered by factors such as fabric wrinkles and illumination changes. An improved YOLOv4-Tiny object detection algorithm is proposed to recognize and locate the sewing stitch points. The number and position information of the sewing stitch points yielded by the model are used for sewing quality detection.

Method The convolutional attention mechanism improved by SoftPool was introduced in YOLOv4-Tiny to enhance the object detection network’s attention to the features of sewing stitch points, before a Soft-SPPF module composed of SoftPool was introduced in front of YOLO Head to realize the model's utilization of multi-scale features in prediction. The improved algorithm was used to calculate the number and coordinate information of all sewing stitch points in the sewing stitch image, together with the density and uniformity of stitch points.

Results The improved YOLOv4-Tiny algorithm and other object detection models were trained by data augmentation on a self-built sewing dataset (Fig. 2) and converge after 150 epochs of training (Fig. 7), and were tested on a test set. By comparing the improved YOLOv4-Tiny algorithm, MobileNet-SSD, YOLOv5s and the original YOLOv4 algorithm, the detection performance in sewing stitch points of the improved method was notably improved. The improved YOLO-Tiny algorithm achieved a mean average accuracy of 85.50% and a detection time of 15.9 ms (Tab. 3), which is more suitable than the original algorithm for sewing stitch detection. In addition (Fig. 8 and Tab. 2), the improved YOLOv4-Tiny algorithm was found to obtain accurately information on the number and coordinates of sewing stitch points in the images of the five sewing stitch types, and density and uniformity were obtained by calculating the number of stitch points per 10 cm and the average relative error between the distance of all adjacent stitch points and the average pixel distance. The quality results obtained from the detection were all within 0.6 stitches per 10 cm difference between density and manual detection, while the maximum difference in uniformity is 1.21%, proving the feasibility of this method. Finally, ablation experiments were carried out and the results are shown in (Tab. 5). The introduction of other modules significantly improved the detection performance of the YOLOv4-Tiny algorithm, and there was no significant decrease in the detection speed. The Soft-CBAM and Soft-SPPF after the introduction of SoftPool improved the detection performance of YOLOv4-Tiny more significantly than the original CBAM and SPPF, and the difference between the detection speed and the original modules was within 0.5 ms, which is a much better performance.

Conclusion The improved algorithm can achieve stitch point recognition despite interference from factors such as fabric folds and accurately obtain location information, which is suitable for sewing stitch detection scenarios. The proposed density and uniformity calculation method is verified by comparing with the manual inspection results, which can meet the actual inspection accuracy requirements and is significantly faster compared with the manual sizing inspection, which can effectively improve the production efficiency of the garment industry.

Key words: clothing production, sewing stitch, quality detection, YOLOv4-Tiny, convolutional attention mechanism, rapid spatial pyramid pooling, clothing quality