Keycap Defect Detection: Simplifying Deep Learning Tasks with Traditional Image Processing

In the manufacturing process of small injection molded products, such as keycaps, defects are inevitable due to potential issues in manufacturing techniques and processes. Defect detection plays a crucial role in ensuring the quality of the final products. However, existing defect detection solutions suffer from low efficiency and accuracy, becoming obstacles to the full automation of production lines in enterprises. In production lines, high-resolution industrial cameras are often required to capture defects, making it challenging to quickly and accurately locate and identify small defects in large-sized images. This paper proposes a method based on the overall layout characteristics of the images and traditional image processing techniques to simplify the final detection task as much as possible. The proposed process involves segmenting a single 5472*3648-sized image into 84 300*300-sized images through operations such as mask matching and edge detection, while eliminating redundant background and reducing task complexity. Finally, deep learning models are utilized to achieve recognition and localization of small defects. Experimental results demonstrate significant improvements in both speed and accuracy compared to directly processing high-resolution images. The proposed method achieves a mean average precision (mAP) of 89.9% using the YOLO-v3 (You only look once) detection model.