Fresco Printing Defect Detection on Cylindrical Yarn Bobbins with Deep Learning and a New Image Dataset

Abstract

The increased pace of production lines in Industry 4.0 has rendered manual quality control processes in the printing and packaging industries insufficient. Cylindrical yarn bobbins with fresco printing are particularly important for production tracking and logistics, especially in yarn winding technologies. Nevertheless, the image distortion caused by perspective and the repetitive pattern designs resulting from the geometry of the cylindrical surface make conventional image-processing-based defect-detection techniques less successful. A quality control system based on deep learning to detect fresco-printing defects on cylindrically shaped paper bobbins is proposed in this work. In the experiment, a new data set comprising 1900 distinct pictures in a real production setting was collected, and the total number of data points was augmented to 3100 through data augmentation methods. MobileNetV2, InceptionV3, DenseNet121, VGG16, and ResNet50 were trained performed using transfer learning, and their performance was analyzed and compared. The experimental findings indicate that the MobileNetV2 model is the most successful architecture, achieving 98.32% accuracy and 99.77% recall. The results show that lightweight CNN models can be used to implement industrial-quality control systems due to their low hardware requirements and high accuracy.