Identification and localization of veneer knot defects based on parallel structure fusion approach

Abstract

Veneer knots are the main indicator of plywood quality. Existing veneer knot identification algorithms have a high identification accuracy rate of over 90%. However, the convolutional neural network (CNN) model is complex and requires laborious data labeling. The localization algorithm produces veneer knot bounding boxes, except for the Mask Region-based CNN (Mask R-CNN) model, which is not accurate and has error transmission. Additionally, the calculation of defect size (area and diameter) has not been addressed. This paper proposes a parallel structured fusion algorithm. One branch employs a classical simple CNN, specifically the Inception V3 network, to identify veneer knot defects. The other branch proposes an improved K-means clustering algorithm to localize veneer knot defects. After identification and localization are achieved, the actual area of the defect is calculated. The proposed method for identifying veneer knot defects has an accuracy rate of 99.61%. The size accuracy localization rate is 94%, with an under-sized localization rate of 2%, an over-sized localization rate of 3%, and the knot localization error rate is 1%. Additionally, the algorithm also calculates the area and diameter of the knot, with a mean absolute error of the diameter of 3.23 mm. This paper presents an algorithm with low complexity, fast speed, high accuracy, and no error transmission, making it suitable for identifying and localizing other defects.