This study presents a comprehensive multimodal dataset and an integrated analysis framework aimed at predicting key structural properties of patterned knitted fabrics using advanced image-based techniques. High-resolution microscope images were collected from forty-five fabric groups, each exhibiting different porosity levels, structural densities, and pattern configurations. For each fabric sample, twenty color images were captured under controlled lighting conditions, and 12 class labels were assigned based on expert-validated porosity categories. The dataset reflects real-world variability due to differences in yarn density, stitch geometry, and color–texture interactions, making the classification task inherently challenging. To evaluate the discriminative power of deep learning models on this fabric dataset, three widely used convolutional neural network architectures—VGG19, EfficientNet-B3, and DenseNet-121—were trained and compared using k-fold cross-validation. Class imbalance was addressed through the use of Focal Loss, and performance was assessed using accuracy, confusion matrices, and fold-averaged metrics. Among the tested architectures, VGG19 achieved the highest overall accuracy, while DenseNet-121 showed balanced performance across several porosity categories. Overall, VGG19 achieved the highest cross-validation accuracy of 87.42% and a final test accuracy of 79.01% demonstrating robust predictive performance for estimating the structural properties of patterned knitted fabrics.
Mehmet Merkepçi (Fri,) studied this question.