A One‐Stage Deep Learning Model for Industrial Defect Detection

Abstract

Industrial defect detection is a hot topic in the field of computer vision and industry. Industrial defects are diverse and complex, and well-known machine learning based methods can often not effectively extract features of industrial defects and achieve good detection results. To address the above problems, this paper introduces a deep learning model for industrial defect detection. First, a two-branch decoupled head, which can facilitate model training through separating the prediction of category and regression is designed. Also, two inverted bottleneck structures are designed to enhance the ability of the model to extract features. Moreover, an attention-enhanced feature fusion (AEFF) module is designed and integrated into the neck network to achieve effective feature fusion. Extensive experiments are conducted on three public datasets, namely the DeepPCB dataset, NEU-DET dataset, and NRSD-MN dataset. The obtained results demonstrate that the proposed model achieves competitive results compared to the state-of-the-art methods. The proposed model achieves mAP@0.5:0.95, 71.78%, 36.04%, and 48.69% on the PCB dataset, NEU-DET dataset, and the NRSD-MN dataset, respectively.