Rapid reconstruction-free classification from Lissajous sparse sampling using deep learning

Non-raster laser scanning techniques such as Lissajous scanning provide mechanically efficient alternatives to conventional raster scanning, but typically require computationally expensive image reconstruction before automated analysis. This reconstruction step offsets many of the speed and efficiency advantages of sparse scanning and is a bottleneck in real-time imaging systems. Here, we demonstrate that accurate object recognition can be performed directly from one-dimensional Lissajous intensity measurements without reconstructing an intermediate image. In our proof-of-concept work, optimized Lissajous scan trajectories were used to sparsely sample MNIST handwritten digits, generating one-dimensional signal representations containing 50, 100, 200, or 500 measurements. These signals were classified using one-dimensional convolutional neural networks (1D-CNN), long short-term memory (LSTM) networks, and k-nearest neighbors (KNN), and compared with conventional reconstruction-based image classification pipelines using both pre-trained and retrained convolutional neural networks. The proposed 1D-CNN achieved 98.60% classification accuracy using only 500 sampled measurements while classifying 10,000 test images in 0.68 s. Reconstruction-based image classifiers achieved slightly higher peak accuracy (99.34%) but required more than 150 s to reconstruct and classify the same dataset. Across all evaluated models, direct signal-domain classification reduced end-to-end computational cost by more than two orders of magnitude while maintaining comparable or slightly diminished classification performance. These findings demonstrate that image reconstruction is not a prerequisite for accurate object recognition from sparsely sampled Lissajous measurements and establish reconstruction-free deep learning as an efficient framework for intelligent laser-scanning systems. The proposed approach has potential applications in LiDAR, optical coherence tomography, endoscopic imaging, image-guided surgery, and other computational imaging systems where rapid inference and reduced computational overhead are critical.