Improved Small Defect Detection on Ceramic Surfaces with YOLOv8

In contemporary engineering construction, ceramics have emerged as a pivotal material, presenting significant challenges in defect detection due to intricate surface patterns and the diminutive nature of flaws. The imperative to augment the precision of ceramic defect identification is twofold: enhancing product quality and elevating manufacturing throughput. This study introduces an advanced method for the identification of surface imperfections in ceramics, utilizing an enhanced YOLOv8 algorithm. Initially, to accommodate the minimal size of ceramic surface defects, images are partitioned into smaller, uniformly-sized segments during the training phase. Subsequently, in the detection phase, a sliding window technique is employed for flaw identification. Observational data analysis revealed that defects predominantly manifest as small, with a minority being medium-sized. To refine detection fidelity, the incorporation of two supplementary detection heads specifically for diminutive targets was enacted. Additionally, to effectively discriminate between complex surface patterns and genuine defects, an innovative selective attention module, dubbed LSKnet, was integrated into the YOLOv8 architecture. Comparative analysis of various Intersection Over Union (IOU) metrics led to the selection of the DIOU metric for this model, aiming to enhance precision further. An experimental endeavor to excise the larger detection head was undertaken to potentially boost model training velocity and accuracy. This methodology was empirically validated on the Tianchi ceramic tile defect dataset, culminating in the improved model achieving an average precision rate of 69.1%, marking an 8.1% accuracy enhancement over the baseline YOLOv8 model following image segmentation.