Single-pixel imaging for edge images using deep neural networks.

Edge images are often used in computer vision, cellular morphology, and surveillance cameras, and are sufficient to identify the type of object. Single-pixel imaging (SPI) is a promising technique for wide-wavelength, low-light-level measurements. Conventional SPI-based edge-enhanced techniques have used shifting illumination patterns; however, this increases the number of the illumination patterns. We propose two deep neural networks to obtain SPI-based edge images without shifting illumination patterns. The first network is an end-to-end mapping between the measured intensities and entire edge image. The latter comprises two path convolutional layers for restoring horizontal and vertical edges individually; subsequently, both edges are combined to obtain full edge reconstructions, such as in the Sobel filter.

Real-time single-pixel imaging using a system on a chip field-programmable gate array.

Unlike conventional imaging, the single-pixel imaging technique uses a single-element detector, which enables high sensitivity, broad wavelength, and noise robustness imaging. However, it has several challenges, particularly requiring extensive computations for image reconstruction with high image quality. Therefore, high-performance computers are required for real-time reconstruction with higher image quality. In this study, we developed a compact dedicated computer for single-pixel imaging using a system on a chip field-programmable gate array (FPGA), which enables real-time reconstruction at 40 frames per second with an image size of 128 × 128 pixels. An FPGA circuit was implemented with the proposed reconstruction algorithm to obtain higher image quality by introducing encoding mask pattern optimization. The dedicated computer can accelerate the reconstruction 10 times faster than a recent CPU. Because it is very compact compared with typical computers, it can expand the application of single-pixel imaging to the Internet of Things and outdoor applications.

Real-valued diffraction calculations for computational holography [Invited].

Computational holography, encompassing computer-generated holograms and digital holography, utilizes diffraction calculations based on complex-valued operations and complex Fourier transforms. However, for some holographic applications, only real-valued holograms or real-valued diffracted results are required. This study proposes a real-valued diffraction calculation that does not require any complex-valued operation. Instead of complex-valued Fourier transforms, we employ a pure real-valued transform. Among the several real-valued transformations that have been proposed, we employ the Hartley transformation. However, our proposed method is not limited to this transformation, as other real-valued transformations can be utilized.

Mitigating ringing artifacts in diffraction calculations using average subtractions.

Fourier transform-based diffraction calculations are essential for computational optics. However, the diffraction calculations can be corrupted by the introduction of strong ringing artifacts due to the introduction of zero-padding to avoid circular convolution or to control the sampling intervals. We propose a simple de-ringing method using average subtractions for application to on-axis and off-axis diffraction calculations. To verify the effectiveness of the proposed method, we compared the diffracted fields obtained using zero-padding, a flat-top window method, a mirror expansion method, and the whole and border average subtractions proposed. Furthermore, we confirmed the effectiveness of the proposed method for hologram calculations using double phase encoding and image reconstructions of inline digital holography.

Single-pixel imaging using a recurrent neural network combined with convolutional layers.

Single-pixel imaging allows for high-speed imaging, miniaturization of optical systems, and imaging over a broad wavelength range, which is difficult by conventional imaging sensors, such as pixel arrays. However, a challenge in single-pixel imaging is low image quality in the presence of undersampling. Deep learning is an effective method for solving this challenge; however, a large amount of memory is required for the internal parameters. In this study, we propose single-pixel imaging based on a recurrent neural network. The proposed approach succeeds in reducing the internal parameters, reconstructing images with higher quality, and showing robustness to noise.

Dedicated processor for hologram calculation using sparse Fourier bases.

Recently, a calculation method involving sparse point spread functions in the short-time Fourier transform (STFT) domain was proposed. In this paper, a dedicated processor using the STFT algorithm is described, which is implemented on a field-programmable gate array. All the operations in this algorithm are implemented using fixed-point arithmetic. Since this algorithm includes a trigonometric function and an error function, lookup tables (LUTs) are utilized to reduce the calculation costs. We have devised a dedicated circuit architecture that allows parallel operations. In addition, a central processing unit could generate holograms using the STFT-based algorithm with fixed-point arithmetic and LUTs at a higher speed than the generation using floating-point arithmetic.

Phase retrieval using axial diffraction patterns and a ptychographic iterative engine.

We propose a phase retrieval method using axial diffraction patterns under planar and spherical wave illuminations. The proposed method uses a ptychographic iterative engine (PIE) for the phase retrieval algorithm. The proposed approach uses multiple diffraction patterns. Thus, adjusting the alignment of each diffraction pattern is mandatory, and we propose a method to adjust the alignment. In addition, a random selection of the measured diffraction patterns is used to further accelerate the convergence of the PIE-based optimization. To confirm the effectiveness of the proposed method, we compare the conventional and proposed methods using a simulation and optical experiments.

Dynamic-range compression scheme for digital hologram using a deep neural network.

This Letter aims to propose a dynamic-range compression and decompression scheme for digital holograms that uses a deep neural network (DNN). The proposed scheme uses simple thresholding to compress the dynamic range of holograms with 8-bit gradation to binary holograms. Although this can decrease the amount of data by one-eighth, the binarization strongly degrades the image quality of the reconstructed images. The proposed scheme uses a DNN to predict the original gradation holograms from the binary holograms, and the error-diffusion algorithm of the binarization process contributes significantly to training the DNN. The performance of the scheme exceeds that of modern compression techniques such as JPEG 2000 and high-efficiency video coding.