关键词: 异性纤维检测, 棉花, 注意力机制, 残差结构, 深度可分离卷积, 网络模型

Abstract:

Objective Aiming at the low detection accuracy of foreign fibers in cotton, and the poor detection effect caused by the hidden or corner position of foreign fibers, the author propose cotton foreign fibers detection algorithm based on residual structure. Its design is based on the characteristics of the different colors and shapes of foreign fibers in cotton, as well as the different shades and light layers of foreign fibers.
Method For the detection target, an online detection scheme for cotton foreign fibers was proposed. A mixed dataset of deep and shallow layers was constructed based on the color, texture, position, and other characteristics of foreign fibers. Building upon this, a foreign fiber detection network model algorithm with a residual structure was designed, addressing issues such as low accuracy in existing detection algorithms and the presence of hidden or corner-positioned foreign fibers.
Results Ultimately, the dataset constructed in this paper includes five common types of foreign fibers, which are plastic film, plastic rope, polypropylene thread, polyester thread, and hair strand. In order to enrich the dataset and increase the validity of the experiment, the dataset also contains samples with no foreign fibers. Comparative experiments are carried out by using the newly developed detection algorithm and other classical algorithms. In the constructed dataset, the experimental results show that the detection accuracy of foreign fibers and plastic film are as high as 92.81% and 95.76%, respectively, while the identification accuracy of hair, polypropylene thread and polyester wire is low (Tab. 1). Plastic film and plastic rope, which are larger than hair strand, are easier to detect. In industry, there are foreign fibers such as hair, polypropylene thread, polyester thread, etc. in the deep and shallow layers of cotton, and the characteristics of these deep foreign fibers are not obvious. Especially the targets of smaller hair strands, After a series of convolutional extraction features. The deep information is more abstracted, so there is less semantic information of such small target objects. In this paper, the attention mechanism is introduced into the residual structure, and different weights are given to the feature map to enhance the representation ability of the key feature of foreign fibers. Experimental results showed that the algorithm reported in this paper achieved remarkable success on deep and shallow datasets. The average detection accuracy of the deep and shallow dataset was 88.48% (Tab. 3). In addition, the algorithm performed well when processing a single image, with an average detection speed of only 0.019 s per image. In comparison to classical algorithms such as GoogleNet, MobileNetV1, MobileNetV2 and EfficientNet, the algorithm improved accuracy by 10.85%, 3.32%, 26.47% and 26.96%, respectively. Compared to MobileNetV2 and EfficientNet, the algorithm tested a single image with shorter running times and higher accuracy. Hence, the algorithm offered a balance between high accuracy and moderate detection speed in foreign fiber detection, catering to the real-time requirements of the industry. It addressed the issues of low accuracy in existing detection algorithms and the presence of hidden or corner-positioned foreign fibers. Furthermore, it introduced a novel approach to foreign fiber detection in cotton.
Conclusion In this paper, the algorithm not only achieves a high detection effect on shallow cotton foreign fibers dataset, but also achieves a good results on the mixed dataset of deep and shallow layers cotton foreign fiber. However, for the small objects in the deep depth of the actual industry, the method in this paper will also have false detection and missed detection. In the future, the network structure and network parameters will be optimized to improve the real-time detection of foreign fibers in cotton while maintaining a high detection accuracy.

Key words: foreign fiber detection, cotton, attention mechanism, residual structure, depth wise separable convolution, network model