loading page

Wooden Texture Anomaly Detection: Leveraging Convolutional Neural Networks and Image Gradients
  • Hussin Ragb,
  • Varshini Nagabooshanam
Hussin Ragb
Electrical and Computer Engineering Department Computer Information System Department, Christian Brothers University Christian Brothers University

Corresponding Author:[email protected]

Author Profile
Varshini Nagabooshanam
Electrical and Computer Engineering Department Computer Information System Department, Christian Brothers University Christian Brothers University

Abstract

Quality assurance in manufacturing often involves manual interventions; however, the integration of machine learning and artificial intelligence can facilitate automation in quality inspection. Industrial process automation significantly contributes to the success of large-scale industries by improving operational efficiency and reallocating resources towards customer satisfaction, thereby enhancing overall business performance. This paper introduces a Convolutional Neural Network (CNN) model based on transfer learning for identifying surface irregularities on finished wooden objects, including stains, porosity, color mismatches, cracks, and knots. Unlike traditional methods that utilize intensity or colored images, our model employs image gradients based on the Canny operator. Edge detection is applied to raw colored images during the image augmentation process to accentuate structural changes on the objects. The rationale behind developing the model on edge-detected images is to reduce noise by eliminating natural patterns that could be misinterpreted as surface irregularities. Deep Convolutional Neural Network models are trained on these edge-detected images. The developed model is assessed using an out-ofsample dataset, with the F1-Score and Accuracy as the primary evaluation metrics. Initially, the model faced challenges in out-of-sample validation due to a low volume of defective images. However, post-performing class balancing, the model exhibited exceptional performance, highlighting its effectiveness in wooden surface irregularity identification.
09 Apr 2024Submitted to TechRxiv
12 Apr 2024Published in TechRxiv