A Comparative Analysis on a Limited Image Dataset for Accurately Detecting Improperly Polished Surfaces for Industrial Applications

The objective of this study is to identify the optimal object detection architecture for training on a specific type of defect detection, namely incorrectly polished surfaces on aluminium elements. In order to facilitate a meaningful comparison of the various architectures, a maximum training time of approximately one hour was established for each architecture. Using the Darknet framework and a specific dataset, five architectures were compared (for the time being). The parameters of the various architectures, including network size, number of batches, and so forth, were modified according to a well-defined and systematic procedure. The preliminary findings indicate that the YOLOv4-tiny network exhibits superior training performance on this dataset, rendering it an optimal choice for industrial applications. This research provides support to small and medium-sized enterprises (SMEs) by identifying effective object detection architectures for quality control and highlighting avenues for advancing AI-driven defect detection in manufacturing.