Efficient target recognition in ghost imaging via optimized 1D convolutional networks at a low sampling ratio

The efficient use of a convolutional neural network (CNN) for object recognition in both conventional imaging and computational ghost imaging (CGI) modalities is demonstrated in this work. The model, which is trained on the Fashion-MNIST dataset, exhibits 92.91% accuracy in image-based classification; class uniqueness has a strong correlation with the performance. More importantly, the model achieves an enhanced accuracy of 97.55% in an image-free CGI framework using only compressed bucket measurements at an 80% sampling ratio. This demonstrates how it can reliably extract features directly from highly subsampled, one-dimensional data, allowing for effective, low-bandwidth optical sensing. We also investigate how resilient the architecture is to noise, and we find that a noise-augmented training approach can significantly reduce performance deterioration in high-noise situations. Thus, the suggested CNN architecture bridges the gap between conventional image processing and applied computational optics by offering a flexible and useful solution for quick, noise-resilient object recognition.