Research on Fabric Defect Detection Based on Multi-branch Residual Network

Abstract

Aiming at the problem that traditional object detection models have low recognition accuracy for small and medium-sized defects. Based on the original residual module, this paper adds a new convolution branch that dynamically adjusts the size of the receptive field with the number of network layers, and then replaces the residual module in the Hourglass-54 down-sampling stage, and proposes a new backbone network: Hourglass -MRB. The experimental results show that the Corernet-Saccade model using Hourglass-MRB improves the recognition accuracy of small and medium-sized fabric defects by 5.8% and 5.6%. The overall recognition accuracy of the system reaches 81.5%. Theoretically,the speed of fabric defect detection reaches 110 m/min. This article provides more effective support for advancing the internationalization of textile quality assessment.