An Improved Neural Network Model Based on DenseNet for Fabric Texture Recognition

Area of Science:

• Computer Science
• Materials Science
• Textile Engineering

Background:

• Automated knitted garment production requires accurate fabric texture identification for quality control.
• Manual methods are inefficient and subjective; current machine learning approaches often rely on time-consuming manual feature extraction.
• Existing methods face limitations in accuracy and efficiency for fabric texture recognition.

Purpose of the Study:

• To develop a novel deep learning model for automated fabric texture recognition in knitted garments.
• To address the limitations of manual feature extraction and improve recognition accuracy.
• To introduce a new dataset for training and evaluating fabric texture recognition models.

Main Methods:

• Introduced the Differentiated Leaning Weighted DenseNet (DLW-DenseNet), an enhanced DenseNet architecture.
• Incorporated a learnable weight mechanism with channel attention to optimize feature selection and reduce redundancy.
• Implemented a differentiated learning strategy with distinct learning rates for continuous channel selection and model pruning.
• Constructed a new knitted fabric dataset named KF9.

Main Results:

• The DLW-DenseNet model demonstrated superior performance in fabric texture recognition.
• Achieved a five percentage point increase in recognition accuracy compared to an improved ResNet-based network.
• Significantly outperformed other representative methods on the KF9 dataset.

Conclusions:

• DLW-DenseNet effectively enhances automated fabric texture recognition in knitted garments.
• The proposed model offers improved accuracy and efficiency over existing methods.
• The KF9 dataset provides a valuable resource for future research in fabric texture recognition.