Single-pixel neural network object classification of sub-Nyquist ghost imaging

Abstract

A single-pixel neural network object classification scenario in the sub-Nyquist ghost imaging system is proposed. Based on the neural network, objects are classified directly by bucket measurements without reconstructing images. Classification accuracy can still be maintained at 94.23% even with only 16 measurements (less than the Nyquist limit of 1.56%). A parallel computing scheme is applied in data processing to reduce the object acquisition time significantly. Random patterns are used as illumination patterns to illuminate objects. The proposed method performs much better than existing methods for both binary and grayscale images in the sub-Nyquist condition, which is also robust to environment noise turbulence. Benefiting from advantages of ghost imaging, it may find applications for target recognition in the fields of remote sensing, military defense, and so on.

Image ciphertexts classification method based on ghost imaging and intraclass-interclass difference

“…Compared with other ciphertext classification methods based on ghost imaging encryption, such as He et al achieved 98% classification accuracy with deep learning [18], and Cao et al achieved 94.23% classification accuracy with neural network [19]. Regardless of the number of samples to be classified, pre-training process is required and the risk of…”

“…Besides, unlike the complex amplitude ciphertexts obtained by most optical encryption schemes, the ciphertext obtained by the CGI based optical encryption process is a set of real numbers, which are highly conducive to storage and transmission. Actually, there have been some researchers who have carried out preliminary research work from the perspective of target classification for ciphertext images, such as He et al proposed a handwritten digit recognition method based on ghost imaging with deep learning [18], and Cao et al proposed a single-pixel neural network object classification method of sub-Nyquist ghost imaging, etc [19]. These methods are at the expense of consuming large computational resources, with the help of machine learning and neural network technology, tens of thousands of ciphertext images are firstly trained to learn the features and then classified, which is able to obtain a high matching ratio of plaintext-ciphertext pairs with low sampling rate.…”

Ghost imaging object recognition based on self-attention mechanism network

“…6 In recent years, researchers also presented a neural network that can recognize target objects without the need for correlated reconstruction, using datasets such as MNIST and Fashion. 7,8 This network model directly incorporates the feature data detected by the bucket detector into a fully connected neural ARTICLE pubs.aip.org/aip/adv network for training, reducing the number of detections and providing a feasible solution for the problem of insufficient feature detail information. However, the recognition accuracy and computational efficiency of these models are still relatively low, and the models are prone to overfitting.…”

Time‐Efficient Object Recognition in Quantum Ghost Imaging

“…[ 12 ] Computationally, intensive attempts to improve imaging speed focused on data acquisition methods, [ 13,14 ] data processing methods using compressive sensing, [ 15–17 ] compressive sensing with generative adversarial networks, [ 18 ] and parallel computing. [ 19 ] Fast image reconstruction algorithms were also implemented as a means to speed up reconstruction, for example, the fast Walsh–Hadamard transform, [ 20 ] and by using logarithmic and exponential ghost‐imaging reconstruction algorithms. [ 21 ] Deep learning has been used to increase image reconstruction quality, [ 22 ] to improve image quality by denoising mechanisms, [ 23–25 ] and recently to reconstruct images at impractical to measure image resolutions.…”