A Multitask Feature Fusion Network for Woven Fabric Density Analysis

Abstract

The density analysis of woven fabrics is a critical part of quality control in textile production. Traditional image-processing-based methods for density analysis of woven fabrics require complicated manual feature design and lack adaptability to different weaving patterns. To address these problems, we propose a woven fabric density analysis method based on small object detection and rule-based post-processing. First, we capture high-resolution images of woven fabrics using macro-microscopic camera equipment, and then construct a woven fabric microscopic image dataset for our study through pre-processing and data augmentation. Next, we propose a multitask feature fusion network (MTF-Net), a small object detection network, to detect the float-points of warp and weft yarns. The detection ability of the model is improved by the cooperation of a reconstruction branch network, a pixel-level branch network, and an object-level branch network. Additionally, we introduce a feature rotation selection module (FRSM) to solve the problem of yarns with small angle rotations. We finally propose a rule-based post-processing method to complete the density analysis of woven fabrics. The experimental results demonstrate that the proposed method is effective and achieves higher accuracy than the popular object detection methods for density analysis on the constructed woven fabric dataset.

Keywords

• Multitask learning
• small object detection
• woven fabric density