Noise Reduction Method for One-Dimensional Signal of Single-Pixel Imaging System in Low-Light Environment

Abstract

The present invention discloses a noise reduction method for one-dimensional signals of a single-pixel imaging system in a low-light environment. A noise reduction convolutional neural network is trained using simulated data to directly perform noise reduction processing on one-dimensional bucket detector signals: First, the properties of noise in a low-light environment are studied to generate a real noise model. A series of simulated noisy signals are generated using this noise model. The simulated noisy signals and simulated noise-free signals are used to train the noise reduction convolutional neural network, and real experimental data is used for testing. The present invention trains the network with simulated data, reducing the cost of data acquisition; it can quickly reduce the noise of noisy signals obtained by short-integration sampling to obtain a reconstructed image with better quality, while improving the imaging efficiency; and it promotes the development of deep learning in the field of single-pixel imaging in a low-light environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and specifically relates to a noise reduction method for one-dimensional signals of a single-pixel imaging system in a low-light environment. Technical Background

[0002] Single-pixel imaging is a method of reconstructing images using correlation measurements. As an emerging imaging technology, it has developed rapidly. Single-pixel imaging technology does not require the use of any array detectors for detection, which gives it the potential to solve some challenges that traditional imaging cannot handle. For example, single-pixel cameras allow people to build a low-cost imaging system that can work at specific wavelengths where array detectors are expensive. In the past decade, single-pixel imaging technology has had many demonstrations in hyperspectral imaging, radar imaging, 3D imaging, and real-time imaging.

[0003] In recent years, researchers have combined deep learning with single-pixel imaging systems to solve some problems in single-pixel imaging and obtained better imaging effects than traditional single-pixel imaging. For example, Lyu proposed using deep learning ghost imaging to construct the GIDL framework, demonstrating that GIDL shows better performance than other methods when the number of measurements is small; Tomoyoshi Shimobab et al. used the U-net network, trained the network with 15,000 128*128 differential ghost imaging reconstructed images and the Calceth-256 dataset as the training set, and used images not in the training set as the validation set, and compared with bilateral filter denoising, achieving better results. Rizvi et al. proposed a fast image reconstruction framework called "DeepGhost", which uses a deep convolutional autoencoder network to achieve real-time imaging at a very low sampling rate (10%-20%).

[0004] In single-pixel imaging, there is always a conflict between imaging efficiency and image quality. In a low-light environment, a long integration time improves image quality but severely reduces imaging efficiency. On the contrary, a short integration time improves imaging efficiency at the expense of image quality. Although the combination of these deep learning and single-pixel imaging has greatly promoted the development of single-pixel imaging, for example, convolutional neural networks have been widely used in image denoising, there is little research on the direct denoising of one-dimensional bucket signals. After denoising the one-dimensional bucket detector signal, a two-dimensional original image better than traditional denoising algorithms can be obtained, bringing great advantages to subsequent image processing. In single-pixel imaging, there is always a conflict between imaging efficiency and image quality. In a low-light environment, a long integration time improves image quality but severely reduces imaging efficiency. On the contrary, a short integration time improves imaging efficiency at the expense of image quality.

[0005] In addition, since the data acquisition of single-pixel imaging still requires a large amount of time, the group of Guohai Situ proposed a method of training a network with simulated data (F. Wang, H. Wang, H. Wang, G. Li, and G. Situ, “Learning from simulation: An end-to-end deep-learning approach for computational ghost imaging,” Opt. express 27, 25560–25572 (2019)), which significantly reduces the cost of data acquisition. However, in low-light environments, there has been no in-depth research on how to simulate the behavior of noise, analyze the properties of noise, and generate simulated noisy signals. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for denoising one-dimensional signals of a single-pixel imaging system in a low-light environment.

[0007] The technical solution for realizing the present invention is as follows: A method for denoising one-dimensional signals of a single-pixel imaging system in a low-light environment, the steps are as follows:

[0008] First step, obtain a marked image and generate a simulated bucket detector signal: In a high-illumination environment (the illuminance of the detector target surface is about 1 lx), use a passive single-pixel imaging system to perform single-pixel detection on a digital target, and reconstruct a clear target image as the marked image; then generate a target library with multiple positions and postures based on the rotation and translation of the image, simulate the single-pixel imaging process, use the target library as the input target for simulation, and generate a simulated noise-free bucket detector signal, denoted as x sim ;

[0009] Second step, obtain the noise parameters in a low-light environment: In a low-illumination environment (the illuminance of the detector target surface is 10 -3 -10 - 4 lx), use a passive single-pixel imaging system to perform long-integration sampling on the target to obtain a noisy signal Separate a relatively clean signal from the noisy signal and noise Analyze the intensity range of the relatively clean signal and establish a linear relationship with the intensity of the simulated noise-free bucket detector signal x sim Make the intensity range consistent with that of x ; Calculate the parameters of the noise sim : mean μ and standard deviation σ, and find the mean of the parameters and and Pair Perform Gaussian simulation;

[0010] In the third step, obtain the simulated noisy signal for network training to generate a denoising model: Add random noise with the same distribution and the same sim to the simulated noise-free bucket detector signal x and to obtain the simulated noisy signal y sim . Multiple groups of x sim and y sim constitute the training data set, which is put into the DnCNN convolutional network for denoising training to generate a denoising model. In the same low-light environment, perform short integration sampling on the target to obtain the experimental noisy signal and use it for network denoising testing, and perform correlation reconstruction on the denoised signal.

[0011] By using the above three steps, the denoised low-light one-dimensional bucket detector signal and a reconstructed image with better quality can be obtained.

[0012] Compared with the prior art, the present invention has the following remarkable advantages: (1) Directly denoise the one-dimensional bucket detector signal under low light. Compared with the traditional method of denoising images, a clearer original image can be obtained, which is more conducive to subsequent image processing; (2) In a single-pixel imaging system, obtaining high-quality images usually requires long integration sampling. By directly denoising the bucket detector signal obtained by short integration sampling under low light using a denoising network, high-quality images are obtained, greatly improving the imaging efficiency; (3) Using simulated data instead of experimental data to train the network greatly reduces the acquisition cost of training samples; (4) Studied the properties of noise under low light, combined experiments and simulations to obtain a more realistic noise model, and solved the problem of how to add real noise to the training samples of the denoising network.

[0013] The present invention will be further described in detail below with reference to the accompanying drawings. Description of the Drawings

[0014] Figure 1 is a schematic diagram of the passive single-pixel imaging system for verifying the present invention.

[0015] Figure 2 is a flowchart of the algorithm for denoising the one-dimensional bucket detector signal under low light of the present invention.

[0016] Figure 3 is a structural diagram of the denoising network DnCNN for denoising one-dimensional signals of the present invention.

[0017] Figure 4 is the image reconstructed before and after denoising of the digital targets 7 and 9. Detailed Embodiments

[0018] The present invention is a method for denoising one-dimensional signals of a single-pixel imaging system in low-light environments. It uses a convolutional neural network for training, reconstructs the denoised one-dimensional signals, and obtains a two-dimensional original image with higher quality, bringing great advantages to subsequent image processing and improving the imaging efficiency at the same time. First, the properties of noise in low-light environments are studied, a real noise model is generated, a series of simulated noisy signals are generated using this noise model, the simulated noisy signals and simulated noise-free signals are used to train the denoising convolutional neural network, and real experimental data is used for testing. Figure 1 It is a schematic diagram of a passive single-pixel imaging system. The main components are an LED light source, a digital micromirror device, a PIN detector, a data acquisition card, and a computer. Figure 2 It is a flowchart of the algorithm of the present invention. The specific steps are as follows:

[0019] Step 1: Obtain a marked image and generate a simulated bucket detector signal. In a high-illumination environment (the illuminance on the detector target surface is 1 lx), use the passive single-pixel imaging system to perform single-pixel detection on the target, and reconstruct a clear target image as the marked image; rotate and translate the marked image to generate a target library with multiple positions and postures, use the target library as the input target for simulation, simulate the single-pixel imaging process, and generate a simulated noise-free bucket detector signal, which specifically includes:

[0020] Step 1.1: First, use an LED strong light source (illuminance is 1 lx), the target O(x, y) is a digital 0-9 with a resolution of 32×32, and a Hadamard matrix I m (x, y) of size 32×32 is shown on the DMD. The target is reflected by the lens to the DMD and modulated by I m (x, y), where m = 1, 2,... 1024. The reflected light intensity of the target is received by the bucket detector of the single-pixel imaging system. The process can be expressed by the following formula:

[0021] B m =∫I m (x, y)O(x, y)dx dy,

[0022] According to the principle of ghost imaging (GI), the image can be reconstructed through the correlation between intensity fluctuations and the illumination speckle pattern:

[0023] G(x, y)=<B m ·I m (x, y)>-<B m ><I m (x, y)>,

[0024] where <·> represents the ensemble average of M groups of patterns, and G(x, y) represents the reconstructed clear image, denoted as the marked image;

[0025] Step 1.2, perform translation and rotation on G(x,y) through simulation to obtain clear images G′(x,y) with different directions and postures. Simulate the experimental system of single-pixel imaging, and use the same Hadamard pattern I m (x,y) to multiply with G′(x,y) to generate a simulated noiseless bucket detector signal of size 1024×1, denoted as x sim .

[0026] Step 2, obtain the noise parameters in low-light environment: In a low-illumination environment (the illuminance on the detector target surface is 7.07×10 - 4 lx), use the passive single-pixel imaging system to perform long-integration sampling on the target to obtain a noisy signal, and separate the relatively clean signal and noise from the noisy signal; analyze the intensity range of the relatively clean signal and establish a linear relationship with the intensity of the simulated noiseless bucket detector signal x sim . Let make be consistent with the intensity range of x sim ; calculate the parameters of the noise: the mean μ and the standard deviation σ, and calculate the mean and standard deviation of the parameters; perform Gaussian simulation on the noise;

[0027] Step 2.1, use the single-pixel imaging experimental system to reduce the brightness of the LED light source to make the illuminance on the detector target surface 7.07×10 -4 lx, perform long-integration sampling on the target. For each frame of the Hadamard pattern, 90000 data points are sampled, and each data point is denoted as to obtain a noisy signal of size 1024×90000 A certain frame of

[0028] Step 2.2, divide into 500 groups to obtain short-integration noisy signals, The number of data points in each group of is 180 (90000 / 500), and the noisy signal of the jth (j = 1…500) group is denoted as

[0029]

[0030] There are a total of j groups of noisy signals, collectively referred to as

[0031] Step 2.3, add up all the data points of and divide by 500 to obtain a relatively clean signal, which is achieved by the following formula:

[0032]

[0033] A set of 1024 relatively clean signals is obtained for 1024 frames

[0034] Step 2.4: Separate the noise. The noise for each frame of the j-th group is achieved by the following formula:

[0035]

[0036] Then the noise of the j-th group is:

[0037]

[0038] The noises of the j groups are collectively referred to as Calculate the mean μ of the noise for each group (j) and the standard deviation σ (j) , finally, calculate the mean of the means and the mean of the standard deviations, which are obtained by the following formulas:

[0039]

[0040] and That is, the parameters required in low-light environments.

[0041] Step 3: Using the noise parameters obtained in Step 2 and as the basis for adding noise, add random noise with the same and to the simulated noiseless bucket detector signal x sim in the simulation to generate the simulated noisy signal y sim .

[0042] Use the set of multiple groups (x sim , y sim ) as the training set to train the denoising network and generate a denoising model. The network uses the DnCNN deep denoising network. The DnCNN algorithm first trains the residual mapping with the residual learning formula Then obtain x = y - R(y), and combine batch normalization to accelerate the training speed and improve the denoising performance. The mean squared error between the desired residual image data and the estimated residual data from the noise input is used as the loss function, and the loss function can learn the trainable parameters θ in DnCNN:

[0043]

[0044] Here, y i represents the actual noisy data, and xi Represents clean data, Represents N pairs of noisy signals and clean signals.

[0045] Under the same low-light environment, the noisy signals of the experiment are obtained through short integration, and are denoised by the denoising model obtained in the second step to obtain the denoised one-dimensional bucket detector signals, and then correlation reconstruction is performed on them

[0046] The algorithm process is carried out according to Figure 2 Proceed, Figure 4 These are the results before and after the denoising of the algorithm of the present invention, and the reconstructed images of different poses of the digital targets 7 and 9 before and after denoising are given. Figure 4 (a)(c) are the reconstructed images of targets 7 and 9 before denoising in 4 different directions, and the images are almost submerged by noise, Figure 4 (b)(d) are the reconstructed images of targets 7 and 9 after denoising, the targets are clear, the noise is reduced, and it can be seen that the method in the present invention can obtain reconstructed target images with higher quality.

Claims

1. A noise reduction method for one-dimensional signals of a single-pixel imaging system in low-light environments, characterized in that the steps are as follows: Step 1, obtain a marked image and generate an analog bucket detector signal: In a high-illuminance environment, use a passive single-pixel imaging system to perform single-pixel detection on the target, and reconstruct a clear target image as the marked image; Rotate and translate the marked image to generate a target library with multiple positions and postures. Using the target library as the input target for simulation, simulate the single-pixel imaging process to generate the simulated noiseless bucket detector signal x sim ; Step 2, obtain the noise parameters in the low-light environment: In a low-illuminance environment, use a passive single-pixel imaging system to perform long-integration sampling on the target to obtain a noisy signal, and separate the relatively clean signal and noise from the noisy signal; analyze the intensity range of the relatively clean signal, establish a linear relationship with the intensity of the analog noise-free bucket detector signal and obtain a coefficient, multiply the relatively clean signal by this coefficient to make its intensity range consistent with that of the analog noise-free bucket detector signal; calculate the parameters of the noise, and find the mean and standard deviation of the parameters; perform Gaussian simulation on the noise, and the steps are as follows: Step 2.1, in a low-light environment, perform long-integration sampling on the target, sampling P data points for each frame, and each data point is denoted as Obtain a noise-containing signal of size 1024×P One frame of is denoted as Step 2.2, divide into N groups to obtain the noisy signals of short integrals. The number of data points in each group of is S, S = P / N, and the noisy signal of the j-th group is The j groups of noise-containing signals are collectively referred to as Step 2.3, add up all the data points of , and then divide by N to obtain a relatively clean signal, specifically: 1024 frames obtain a relatively clean signal group of 1024×1 Establish a linear relationship with x sim Establish a linear relationship with x will Make Establish a linear relationship with x sim The intensity ranges are consistent; Step 2.4, separate the noise, and the noise of each frame in the jth group is: Then the noise of a total of 1024 frames in the jth group is: The j groups of noises are collectively referred to as Calculate the mean μ of each group of noises (j) and the standard deviation σ (j) , and calculate the mean of the means and the mean of the standard deviations, which are obtained by the following formulas: and i.e., the parameters required in low-light environments; Step 3, obtain an analog noisy signal for network training and generate a noise reduction model: Add random noise with the same distribution and the same mean and standard deviation of parameters to the analog noise-free bucket detector signal to obtain an analog noisy signal; obtain a training data set composed of multiple groups of analog noise-free bucket detector signals and analog noisy signals, put it into the DnCNN convolutional network for noise reduction training to generate a noise reduction model. In the same low-illuminance environment, perform short-integration sampling on the target to obtain an experimental noisy signal and use it for network noise reduction testing, and perform correlation reconstruction on the denoised signal.

2. The noise reduction method for one-dimensional signals of a single-pixel imaging system in low-light environments according to claim 1, characterized in that the steps of obtaining the marked image and generating the analog bucket detector signal are as follows: Step 1.1, under strong light illumination, select a digital target that is reflected by a lens onto the DMD, modulated by a Hadamard matrix pattern, and the reflected light intensity of the target is received by the bucket detector of the single-pixel imaging system. The process is represented by the following formula: B m = ∫I m (x,y)O(x,y)dx dy, where I m (x, y) is a Hadamard matrix and O(x, y) is the target; According to the principle of correlated imaging, reconstruct the image through the correlation between intensity fluctuations and illumination speckle patterns: G(x,y) = <B m ·I m (x,y)> - <B m >>I m (x,y)>, where <·> represents the ensemble average of M groups of patterns, and G(x, y) represents the reconstructed clear image, which is used as the marked image; Step 1.2: Translate and rotate the marked image to obtain clear images with different postures, and form a target library; simulate the experimental system of single-pixel imaging, multiply with the same Hadamard pattern, and generate a one-dimensional simulated noiseless bucket detector signal, denoted as x sim .

3. The noise reduction method for one-dimensional signals of a single-pixel imaging system in low-light environments according to claim 1, characterized in that the network in Step 3 uses a DnCNN deep noise reduction network. The DnCNN network first trains the residual mapping R(y) using the residual learning formula to obtain x = y - R(y), and then combines batch normalization to accelerate the training speed. The mean squared error between the expected residual image data and the estimated residual data from the noise input is used as the loss function, and the loss function learns the trainable parameters θ in the DnCNN: y i represents the actual noisy data, and x i represents the clean data.

Images