Accurate subpixel center determination for overlapping spots

Abstract

Precisely detecting laser beam locations is crucial in maximizing the performance of optical systems in manufacturing and measurement applications. There are numerous methods dealing with a single spot, but identifying overlapping spot centers is still challenging. To address this issue, we present an innovative approach that uses convolutional neural networks and image processing techniques to localize overlapping spot centers. Our method begins by utilizing the convolutional neural network to extract two crucial features: the ratio of minor to major axes and the orientation of the spots. Then, the Euclidean distance transform is employed to identify the approximate centers of the spots, which are the positions with the highest intensity in the transformed images. Finally, the gradient descent algorithm is applied to determine the precise center locations. In addition, noise was added to examine the ability to work in actual systems. The results indicate that our method performs well in noisy environments, accurately pinpointing overlapping spot centers in real-time with a 92% success rate. Furthermore, our approach strikes an excellent balance between accuracy and computational efficiency, making it suitable for use in actual laser manufacturing systems.