Research on Vortex Light Atmospheric Detection Based on Dataset Optimization Algorithm

When vortex beams propagate in space, they inevitably encounter the influence of atmospheric turbulence. Usually, we collect images of vortex beam intensity transmitted through atmospheric turbulence and use them for deep learning to invert turbulence parameters. However, simply feeding all the collected images into the network not only increases the memory usage of the neural network, but also increases the time consumed for network training, which is not conducive to efficient and fast network training and turbulence inversion. Moreover, it can bring certain interference information, resulting in poor accuracy of training and even overfitting. In order to improve the accuracy and efficiency of turbulence inversion, this paper proposes a dataset optimization algorithm based on feature pre extraction. By maximizing the differences in the dataset, the goal is to preserve the maximum difference features while minimizing the presence of misleading features. Before sending images into a convolutional neural network for classification and recognition, we extract the region with the highest variance based on the variance of each group of images, and use the cropped and normalized images as inputs to the network. The results showed that, under the same amount of data and after the same number of iterations of training, the accuracy of the network’s classification and recognition of optimized datasets could reach 79.37% after training. Compared with the accuracy of classification and recognition of unoptimized datasets, it improved by about 13.5%. The time spent on network training was also reduced from 4 hours to 10 minutes, and the convergence speed of the network was greatly accelerated in the early stage of network training. The recognition accuracy of the final validation set is 77.25%. This feature pre extraction method based on maximizing differences provides a new approach for exploring how to optimize turbulence inversion to achieve accurate and efficient fast inversion.