A deep learning-based model proposal for a quality control system

Wood as a raw material is currently used in many industries. Unsightly texture and other defects on the wood make it less visually appealing and impact its durability as a material. Blemishes on the wood reduce its value, and defective material should be identified and removed from production. Since it's difficult to spot defects through visual or manual inspection, an automated system integrated into the production system is a viable option. With the help of advanced digital technology, automated systems for such controls are being developed. In recent years, researchers have been developing quality control applications using image processing-based inspection systems. By means of these technologies, defects in the raw material can be spotted in the early stages of production. Early identification of defects will prevent the production of faulty end products and will reduce labor and material losses in production. In this study, an integrated system is proposed to identify defective material. The proposed system acquires images of the raw material through a special camera installed on the production line. Digital images will be analyzed by deploying deep learning methods, and defective materials will be set apart. Convolutional Neural Network (CNN), a distinction-based method in image processing for detecting defects, was preferred because of its suitability. In addition, different CNN architectures such as ShuffleNet, AlexNet, GoogleNet, and parameters associated with these architectures were tested to identify the most suitable architecture for this problem. In the CNN method, considered in the study for quality control systems to separate out defective wood products, MobileNet, DenseNet, and Inception architectures gave promising results. In addition, image augmentation and image enhancement methods were added to the CNN method experiments, both separately and together, and their effect on performance metrics was examined. In each trial, parameters were modified, and the impact of parameters on the performance metrics was examined. Consequently, CNN architecture was selected with the parameter set giving the best performance.