MSCL-Net: Multispot Centers Localization Network of Optical Aspheric Surface Measurement With Multibeam Angle Sensor

The performance of aspheric optics is important for optical imaging and profiling is required to ensure their performance. When using the self-developed multibeam angle sensor (MBAS) for measurements, the measurement accuracy is reduced by noise and the efficiency is low due to the complexity of the center localization algorithm. Therefore, in this research, a multispot centers localization network (MSCL-Net) is proposed for multispot center localization. The network comprises three components: backbone, FPN-Neck, and detector. First, the backbone and FPN-Neck perform spot feature extraction and feature fusion to obtain high-level semantic information. Second, a localization attention module (LAM) is set in the detector to improve the network’s capability in extracting spot features and suppressing noise. Finally, a localization loss function is introduced to increase the importance of the spot center loss in the latter stages of training, optimizing the spot center localization. The experimental results demonstrate that this method processes a multispot image in 20.3 ms, and spot center localization accuracy to subpixel level. For different radii of curvature optical elements, the method achieves measurement standard deviations (STDs) as low as $0.0041~\mu $ m.