Phase recovery method of depth physical parameter integrated network for lensless microscopic imaging

Abstract

The invention discloses a depth physical parameter integrated network and method for lensless microscopic imaging, and the method comprises the steps: firstly building a physical forward model based on fractional Fourier transform, and quantizing a propagation distance into an adjustable fractional order parameter; secondly, constructing a multi-stage progressive network framework, wherein each stage comprises a physical model fidelity module and an image regularization updating module; a self-adaptive near-end network and a self-adaptive compression-excitation network are innovatively embedded in a regularization module, and real-time self-adaption of network parameters to imaging conditions is achieved; finally, end-to-end single-frame reconstruction is realized, complex amplitude information is directly recovered from a single intensity observation image, and retraining for different imaging parameters is not needed. According to the method, the reconstruction physical reliability, the noise robustness and the cross-scene generalization ability are remarkably improved, meanwhile, the lightweight real-time deployment requirement is met, and the method is suitable for lens-free microscopic imaging scenes such as biological living body observation and portable pathological diagnosis.

Description

Technical Field

[0001] The present invention relates to the field of computational optical imaging and deep learning technology, and in particular to a phase recovery method of a deep physical parameter integrated network for lensless microscopy. Background Art

[0002] Lensless microscopy omits the lens components of traditional microscopes and utilizes computational imaging to reconstruct the complex amplitude information of the sample from the diffraction pattern. This technology offers the advantages of a simple structure, a large field of view, and low cost. However, optical detectors only record the intensity of light waves and lose their phase, resulting in a serious ill-posed inverse problem in the imaging process. Existing methods suffer from three key flaws: First, multi-frame imaging strategies rely on sophisticated hardware. These include multi-sample-sensor distance adjustment and multi-wavelength or multi-angle illumination. While these strategies improve accuracy, they require mechanical displacement platforms or multi-light source systems, significantly increasing complexity and cost. Furthermore, data acquisition is time-consuming and cannot meet the needs of real-time observation. Second, single-frame deep learning methods have inherent limitations. Physics-driven methods (such as embedding Fresnel diffraction models) require manual fine-tuning of hyperparameters, resulting in complex training and poor noise robustness. Data-driven methods (end-to-end black-box mapping) lack interpretability, require retraining when the data distribution changes, and exhibit weak generalization capabilities. Third, integrated network frameworks ignore physical modeling. Although the codec or Transformer structure uniformly handles multiple degradation types, it does not explicitly model the physical mechanism of light wave propagation (such as the quantitative relationship between propagation distance and diffraction effect), resulting in the inability of network parameters to adapt to dynamic imaging conditions (such as noise level and fractional order changes), and severe loss of high-frequency details during long-distance imaging.

[0003] The fundamental technical bottleneck lies in the fact that existing solutions fail to balance physical interpretability, real-time performance, noise robustness, and scenario generalization. Multi-frame strategies sacrifice real-time performance; single-frame methods struggle with physics-driven optimization and lack data-driven generalization; and the integrated network struggles to adapt to complex, degraded scenarios due to the lack of a parameterized propagation distance model and dynamic control mechanisms. The core challenge lies in establishing a quantifiable physical forward model and enabling real-time adaptation of network parameters to imaging conditions, while also meeting the requirements of lightweight deployment. Summary of the Invention

[0004] In view of the defects of the prior art, the present invention provides a phase recovery method of a deep physical parameter integrated network for lensless microscopy.

[0005] In order to achieve the above object of the invention, the technical solution adopted by the present invention is as follows:

[0006] A phase retrieval method for a deep physical parameter integrated network for lensless microscopy imaging comprises the following steps:

[0007] A physical forward model of lensless microscopy based on fractional Fourier transform is established to map the diffraction propagation distance d of the light field into a controllable fractional order p.

[0008] Constructing a progressive deep physical parameter integration network framework, the framework comprises M sequentially connected stages, each stage including: a physical model fidelity module and an image regularization update module;

[0009] Physical model fidelity module: performs a forward fractional Fourier transform based on the fractional order p on the input image, replaces the amplitude with the actual collected observation intensity, and then performs an inverse fractional Fourier transform to output the updated estimated image;

[0010] Image regularization update module: This module performs feature enhancement on the output of the physical model fidelity module, including:

[0011] Adaptive proximal network: dynamically generates affine transformation parameters through fractional order p and noise intensity a to achieve conditional adaptive instance normalization;

[0012] Adaptive compression-excitation network: The weights and biases of the fully connected layer are regulated by the fractional order p and noise intensity a, and channel attention weights are generated for feature weighting.

[0013] A single-frame intensity observation image is input into the network, and the recovered complex amplitude information is directly output without the need for retraining for different fractional orders or noise levels.

[0014] Furthermore, the physical forward model satisfies:

[0015]

[0016] And p increases in the range [0,1];

[0017] Where d represents the diffraction propagation distance and λ represents the wavelength;

[0018] And the intensity I recorded by the detector d The relationship with the sample complex light field U0 is:

[0019]

[0020] Among them, I d (x,y) represents the intensity recorded at the coordinate (x,y), F p represents the fractional Fourier transform operator, U0 represents the complex light field in the sample plane, and x, y represent the spatial coordinates of the sensor plane.

[0021] Furthermore, the adaptive proximal network implements feature adaptation through the following formula:

[0022]

[0023] in, represents the conditional adaptive instance normalization operation of the jth layer in the kth stage, represents the input feature map, represents the channel mean of the feature map, represents the channel standard deviation of the feature map, represents the scaling factor generated dynamically by the fractional order p, represents the translation factor dynamically generated by the fractional order p;

[0024] And the scaling factor and translation factor satisfy:

[0025]

[0026] in, represents the fully connected layer, and C represents the number of feature channels.

[0027] Furthermore, the adaptive compression-excitation network includes:

[0028] Compression operation: compress the feature map into a channel description vector through global average pooling;

[0029] Excitation operation: It consists of two fully connected layers, whose weights and biases are dynamically generated by fractional order p regulation, and output channel weight vectors;

[0030] Fusion operation: Multiply the channel weights by the input features channel by channel, then perform convolution on the output, and add a residual connection to retain the original information.

[0031] Furthermore, the noise intensity a is a Poisson noise parameter, and its value range is a∈[10,50], where a smaller a value indicates a greater noise intensity.

[0032] The present invention also discloses a phase recovery device for lensless microscopic imaging, comprising:

[0033] A physical forward modeling unit configured to establish a light field propagation model based on fractional Fourier transform;

[0034] Progressive processing unit: contains M stages in series, each stage includes:

[0035] Physical model fidelity module: configured to perform fractional Fourier forward / inverse transform and intensity projection;

[0036] Image regularization update module: This module includes an adaptive proximal network and an adaptive compression-excitation network, both of which dynamically control network parameters through fractional order p and noise intensity a;

[0037] Output unit: configured to recover the complex amplitude directly from a single frame of observation image.

[0038] The present invention also discloses a lensless microscopic imaging system, integrating the above-mentioned phase recovery device, wherein:

[0039] The system uses a CMOS sensor to collect diffraction patterns, processes single-frame observation images through the device, and directly outputs complex amplitude reconstruction results of biological samples.

[0040] Compared with the prior art, the advantages of the present invention are:

[0041] 1. Physically interpretable high-precision reconstruction

[0042] Through rigorous physical modeling of light wave propagation, the complex diffraction process is transformed into a controllable fractional-order transformation, ensuring that the reconstruction results strictly conform to the laws of optics. In challenging scenarios such as long propagation distances, the clarity of microstructures such as cell edges and microvessels is significantly improved, effectively addressing the structural distortion issues inherent in traditional black-box models.

[0043] 2. Strong robustness in complex noisy environments

[0044] The innovative approach embeds noise intensity as a dynamic control parameter into the core network module, enabling the reconstruction process to adaptively match the actual noise distribution. Even under strong noise interference, the reconstructed image is free of artifact accumulation and retains detail far exceeding existing mainstream solutions.

[0045] 3. Zero-cost generalization capability across scenarios

[0046] Leveraging a physical parameter-driven adaptive mechanism, the network can be directly applied to new scenarios, such as those with varying propagation distances and unknown noise types, without retraining. Under fractional-order conditions not covered by training, the fluctuation in reconstruction quality is reduced, completely resolving the generalization problem of traditional methods.

[0047] 4. Real-time and efficient embedded deployment

[0048] Using a lightweight model structure and a progressive reconstruction framework, the inference time for a single frame is reduced to milliseconds. Its low parameter count enables real-time dynamic observation in resource-constrained mobile microscopes.

[0049] 5. Multi-parameter collaborative optimization capabilities

[0050] It achieves a breakthrough in unified control of multi-dimensional physical parameters such as propagation distance and noise level, and strikes a balance between imaging speed, accuracy, and anti-interference capabilities in complex tasks such as cell dynamic monitoring, providing a new solution for the study of living biological samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of a phase recovery method for a deep physical parameter integrated network according to an embodiment of the present invention;

[0052] Figure 2 This is the Set12 test set of the embodiment of the present invention;

[0053] Figure 3 This is the Unnatural6 test set diagram of an embodiment of the present invention;

[0054] Figure 4 Graphs showing the reconstruction results of the natural image PatrickStar using fractional Fourier observations corresponding to different fractional orders using six phase recovery methods according to an embodiment of the present invention;

[0055] Figure 5 Graphs showing the reconstruction results of fractional Fourier observations corresponding to different fractional orders of the unnatural image Pollen using six phase recovery methods according to an embodiment of the present invention;

[0056] Figure 6 This is a comparison of the visualization reconstruction performance of the method of the present invention and the all-in-one supervised learning method - AdaIR under Poisson noise interference with an intensity of α=30;

[0057] Figure 7 This is a comparison of the reconstructed samples with a diffraction propagation distance of 1.716 mm according to an embodiment of the present invention; (a) is the observed pattern, and (b) is the restored image;

[0058] Figure 8 This is a comparison diagram of the reconstructed sample with a diffraction propagation distance of 3.912 mm according to an embodiment of the present invention; (a) is the observed pattern, and (b) is the restored image;

[0059] Figure 9 This is a comparison diagram of the reconstructed samples with a diffraction propagation distance of 6.148 mm according to an embodiment of the present invention; (a) is the observed pattern, and (b) is the restored image. DETAILED DESCRIPTION

[0060] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples.

[0061] The present invention provides a phase recovery method for a deep physical parameter integrated network for lensless microscopy imaging, comprising:

[0062] 1. Constructing a physical forward model for lensless microscopy

[0063] In a lensless microscopy system, the biological sample is placed just a few millimeters above the image sensor, and the imaging process does not require the use of optical lenses or other bulky mechanical equipment. According to Maxwell's equations, the propagation process of the sample's complex light field from the initial plane z = 0 to another plane z = d can be approximately expressed by the Fresnel integral:

[0064]

[0065] Where λ and d are two physical parameters, representing the wavelength and propagation distance of light respectively, U represents the light field at different positions, and i is an imaginary number. The final diffraction pattern recorded by the detector is I d (x,y)=|U d (x,y)| 2 , the two-dimensional fractional Fourier transform of the light field distribution U0(x′,y′) can be defined as:

[0066]

[0067] The value range of the fractional order is 0<|p|<2, and the kernel function of the fractional Fourier transform is defined as:

[0068]

[0069] Relating the Fresnel integral (Formula 1) with the fractional Fourier transform (Formula 3) and leveraging its efficient discrete properties yields:

[0070]

[0071] in As the propagation distance d increases, the corresponding fractional order p increases within the range [0, 1]. Leveraging the properties of the fractional Fourier transform and its efficient discrete implementation, the above formula can serve as a universal, accurate, and computationally efficient physical forward model for lensless microscopy. Based on this, this embodiment can construct a synthetic paired sample image / observation pattern dataset for subsequent network training.

[0072] 2. Constructing a Progressive Deep Physics Parameter Integrated Reconstruction Network Framework

[0073] Based on the physical forward model of lensless microscopy established above, the phase retrieval problem is mathematically transformed into a regularized optimization problem by minimizing the following cost function:

[0074]

[0075] in is the reconstructed image, I is the collected diffraction pattern, F p (·) is a physical propagation model based on fractional Fourier transform, is a regularization term, and β is a hyperparameter used to control the weight of the regularization term.

[0076] In order to solve this non-convex nonlinear pathological problem, the present invention proposes a progressive reconstruction framework, which consists of M stages, each stage including two main modules: a physical model fidelity module and an image regularization update module, thereby gradually improving the network performance. Among them, the image regularization update module further includes two sub-modules: an adaptive proximal network and an adaptive compression-excitation network. By introducing controllable physical parameters, a dynamic adjustment mechanism is established to achieve efficient parameterization and dynamic adaptive reconstruction, thereby improving the stability and accuracy of reconstruction. The network structure proposed in the present invention is trained using an end-to-end supervised learning method. After the training is completed, the corresponding original object complex amplitude information can be directly recovered from a single intensity observation image. It has the advantages of compact structure, high recovery accuracy and high inference efficiency.

[0077] 1. For the k-th stage physical model fidelity part, a forward propagation model based on fractional Fourier is constructed, and an analytical intensity projection strategy is used to iteratively optimize the initial reconstructed image. Specifically, the current estimated image is first Perform forward transformation to obtain its fractional domain representation and replace its amplitude with the observed intensity I to retain the phase information; then, perform inverse transformation F -p (·) Return to the image domain and obtain the updated estimated image

[0078]

[0079] in represents the current reconstructed image, represents the Hadamard product, The output image representing the fidelity part of the physical model at the current stage.

[0080] 2. For the image regularization update part of the kth stage, this embodiment uses submodule 1 - adaptive proximal network to achieve feature extraction and enhancement, and uses submodule 2 - adaptive compression-excitation network to further adjust the channel feature response. Controllable physical parameters (such as fractional order and noise intensity, etc.) are introduced in these two key modules to achieve efficient parameterization and dynamic adaptive reconstruction, thereby improving the performance of the network. Represents the output image of the current image regularization update part:

[0081]

[0082] 2.1 Building an Adaptive Proximal Network

[0083] This submodule aims to adaptively extract and enhance features based on physical parameters, thereby more effectively solving the regularization update part of formula (7).

[0084]

[0085] where k = 1, 2, ... M, j = 1, 2, 3, 4. Specifically, represents the output of the k-th stage physical model fidelity part, that is, the input of the k-th stage adaptive proximal network (submodule 1 of the image regularization update part), and represents the output of the k-th stage adaptive proximal network (submodule 1 of the image regularization update part). CAdaIN is a key component of the adaptive proximal network. It normalizes and recalibrates the mean and variance of image features to achieve style transfer and feature alignment of content features:

[0086]

[0087] in, Represents the j-th input feature map of the k-th stage. and Represents the input feature map Here, the fractional order p is used as a controllable variable to dynamically adjust the normalization process. is the learned scaling factor used to adjust the magnitude of the normalized value, is the learned translation factor used to reposition the normalized value. The expression of the above learnable affine parameters is as follows:

[0088]

[0089] in, and represents different fully connected layers that are flexibly adjusted according to p, and C represents the number of channels.

[0090] Considering that actual observation images are typically acquired at different fractional orders, submodule 1 dynamically generates the mean and variance required for normalization based on the input physical parameter p and adjusts the affine transformation process in the feature map accordingly. In other words, this structure can adaptively control the affine parameters in the normalization process based on the amplitude or intensity observation information. Therefore, the proposed adaptive proximal network not only achieves the cross-conditional migration capability between measurement data acquired under different imaging conditions, but also effectively enhances the model's generalization ability for unknown imaging parameter configurations, improving the network's adaptability and robustness in various practical application scenarios.

[0091] 2.2 Building an Adaptive Compression-Excitation Network

[0092] This submodule embeds a special physical parameterized attention mechanism that generates refined channel weights by aggregating key features, thereby improving the adaptability and flexibility of the model. In addition, it can be seamlessly integrated into various convolutional neural networks as a lightweight gating module. The specific steps include the following.

[0093] 2.2.1 Squeeze operation

[0094] This operation aggregates the spatial dimensions of the feature map and compresses it into a single-channel representation. In essence, it extracts and maps the global spatial information into a channel vector to capture the global context features of the image. Output of the k-th stage adaptive proximal network The compression operation is achieved through the following formula:

[0095]

[0096] Here, H and W are the height and width of the input, respectively. This operation aggregates global spatial information into a compact channel representation, providing a basis for the subsequent adaptive channel weighting mechanism, thereby achieving more discriminative feature enhancement and image reconstruction.

[0097] 2.2.2 Excitation operation

[0098] This operation captures channel correlations through a simple gating mechanism and a sigmoid activation function, enabling the network to achieve feature recalibration. In the traditional SENet architecture, this step feeds the channel description vector obtained in the first step into a lightweight fully connected network to learn the importance of each channel, as shown in the following formula.

[0099]

[0100] Among them, the fully connected layer L1 is used to reduce the dimension of the feature vector A, while the fully connected layer L2 generates new weights on the original dimension by combining the significant features after dimensionality reduction to achieve feature reconstruction.

[0101] Although the traditional SENet architecture can improve feature representation capabilities by modeling dependencies between channels, its excitation mechanism relies on static parameter settings, lacks dynamic adaptability to different signal features, and is difficult to achieve fine-tuning. To overcome the above limitations, the present invention proposes an enhanced channel attention structure, called an adaptive compression-excitation network, which introduces controllable physical parameters into the channel excitation mechanism. Specifically, the important parameters in the fully connected layer mentioned above, namely weights and biases, are regulated by introducing fractional order p. Unlike the traditional SENet that uses fixed parameters, the weights and biases in this structure are dynamically generated through multiple new fully connected layers, and its structural dimensions are set to Where C is the number of channels of the input feature and r is the compression ratio

[0102]

[0103] in, and There are two different fully connected layers used to generate the weights of the L1 layer Responsible for generating its bias Similarly, for the L2 layer, its weights and biases are also generated by independent fully connected layers to ensure the controllability and diversity of parameter generation.

[0104] 2.2.3 Fusion operation.

[0105] Finally, each element in the generated weight vector B corresponds to a channel, and different channel features are weighted and scaled by element-by-element multiplication:

[0106]

[0107] Through the channel-by-channel scaling mechanism, the model can perform differentiated processing based on the importance of different channels, thereby enhancing the overall feature representation capability and network performance, and improving reconstruction accuracy. In addition, a residual component is introduced here It is used to preserve the original global context information and perform auxiliary enhancement on the reconstructed features of the main branch to ensure that important structures and details are not lost during the information compression and excitation process.

[0108] The above steps work together to promote high-quality reconstruction of the light field of complex biological samples, and ultimately achieve accurate recovery of the target complex light field.

[0109] The main advantage of the present invention is that the fractional Fourier is used as a flexible and accurate physical forward model to construct a lensless microscopy imaging system, in which physical parameters such as fractional order and noise intensity are used as adjustable parameters to describe different lensless microscopy imaging settings. In addition, the lensless microscopy single-frame phase recovery method based on a deep physical parameter integrated network proposed in the present invention adopts a progressive reconstruction network and generates more refined prior information in each reconstruction stage by introducing an adaptive proximal module and an adaptive compression-excitation module with adjustable parameters. After training, the integrated network structure can be directly applied to single-frame phase recovery tasks under different fractional orders and noise levels without retraining, and has high physical reliability, high-fidelity reconstruction, high computational efficiency, high robustness and high generalization ability in the lensless microscopy imaging system.

[0110] Experimental verification: In order to train the deep physical parameter integration network proposed in this invention, 500 images from the Berkeley Segmentation Dataset (BSD dataset) were selected and each image was divided into 12 256×256 image blocks, thereby constructing a training set containing a total of 6000 high-quality images. At the same time, different fractional Fourier observations were obtained by uniformly sampling fractional orders from the discrete set {0.1, 0.2…0.9}. This sampling strategy limits the fractional order to a predefined range, while retaining rich representation capabilities, it effectively improves the stability of the training process. Taking into account noise interference, Poisson noise ∈ is further superimposed on the training samples, and the noise level α is uniformly extracted from the interval [10,50] to enhance the robustness and generalization ability of the network in complex imaging environments.

[0111] To evaluate the proposed method, this embodiment compares five mainstream phase recovery methods, including three methods based on unsupervised deep learning: prDeep, PhysenNet, DeepMMSE, and two supervised learning methods based on All-in-one architecture: TransWeather, AdaIR. peDeep was tested on Matlab2023a. For the PhysenNet and DeepMMSE methods, the Adam optimizer was used in the training process, and 10,000 rounds were trained with learning rates of 4e-4 and 1e-4, respectively. For the two methods based on All-in-one architecture, this embodiment retains their original network structure and hyperparameter configuration, and retrains and evaluates them on the same training data as the present invention to ensure fairness and consistency of the comparison. In addition, this embodiment adopts the public dataset Set12 ( Figure 2 ) and Unnatural6( Figure 3) to quantitatively evaluate the performance of the algorithm. Twelve widely used natural images and six non-natural images are included, and all images are resized to 256×256.

[0112] Here, this example compares the reconstruction performance of six phase retrieval methods (including the present invention) for different fractional orders and their observations. Table 1 reports the reconstruction results of each method on the test datasets Set12 and Unnatural6, using peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) as evaluation metrics. Although the AdaIR method exhibits relatively good restoration performance at low fractional orders (e.g., p = 0.1) in the Set12 dataset, its architecture relies solely on learning the mapping between the degraded image and the original image, lacking an explicit physical forward model. Therefore, as the fractional order increases (i.e., the propagation distance between the light source and the sensor in a lensless microscopy setup increases), its restoration performance significantly decreases. In particular, when the fractional order approaches 1 (e.g., p = 0.9), the fractional Fourier domain approaches the traditional Fourier domain, further exacerbating imaging degradation and resulting in a significant decrease in reconstruction quality for most methods. In contrast, the deep physical parameter integration framework proposed in this invention maintains robust performance and achieves optimal image reconstruction, thanks to its synergistic integration of physical models and prior learning. In order to more intuitively demonstrate the reconstruction results of each method, this example takes an image from the Set12 test set and the Unnatural6 dataset as examples to show the reconstruction results of the fractional Fourier observations corresponding to these six phase recovery methods at different propagation distances. Figure 4 and Figure 5 shown.

[0113] Table 1 Comparison of reconstruction performance of various phase retrieval methods under different fractional orders on Set12 and Unnatural6 datasets (average peak signal-to-noise ratio: dB / average structural similarity)

[0114]

[0115]

[0116] In addition, in order to comprehensively evaluate the universality of the model, Table 2 shows the reconstruction performance of three supervised neural networks based on the All-in-one architecture in unknown fractional orders (i.e., fractional orders that have not appeared during the training process). The experimental results show that although some methods perform well when the fractional orders are known (i.e., fractional orders seen during the training process), under unknown order conditions, due to their lack of intrinsic modeling capabilities for physical change conditions, their generalization capabilities are insufficient, and the reconstruction performance significantly decreases when facing unknown fractional orders in the testing phase. In contrast, the deep physical parameter integrated network proposed in the present invention achieves adaptive modeling of different propagation conditions by relying on explicit physical forward modeling and controllable parameter mechanisms. When facing unknown fractional orders, it can still maintain stable and excellent reconstruction effects, which fully demonstrates its good robustness and wide generalization capabilities.

[0117] Table 2 Comparison of reconstruction performance of different phase retrieval methods based on all-in-one architecture under unseen fractional orders during training (test datasets: Set12 and Unnatural6; average peak signal-to-noise ratio: dB / average structural similarity)

[0118]

[0119]

[0120] In order to be closer to real imaging conditions and simulate actual application scenarios, Poisson noise is introduced in the experiment, and the noise intensity α is uniformly sampled from the range of [10,50]. In this noise addition method, a smaller α value corresponds to a lower signal-to-noise ratio, which means that the image degradation is more serious. Given that the AdaIR method performs best among the existing methods based on the All-in-one architecture and has been widely regarded as the most advanced model, only this method is selected as a comparison benchmark in this part of the experiment. To ensure the fairness of the experiment, this embodiment adds Poisson noise of the same intensity to the method. As shown in Table 3, regardless of whether the noise intensity is high or low (that is, the signal-to-noise ratio is high or low), the AdaIR method shows high sensitivity under different noise levels, its reconstruction performance is significantly reduced, and its robustness is insufficient. In contrast, the method proposed in the present invention maintains excellent restoration effects under conditions of any fractional order containing noise, fully demonstrating its robustness and stability under complex degradation conditions. Figure 6 The visualization shows the reconstruction results under the condition of α=30.

[0121] Table 3 Comparison of reconstruction performance of the proposed method and the advanced all-in-one supervised learning method - AdaIR under different intensities of Poisson noise (test datasets: Set12 and Unnatural6; average peak signal-to-noise ratio: dB / average structural similarity)

[0122]

[0123] In order to further evaluate the computational efficiency and deployment feasibility of the method of the present invention, Table 4 lists the comparison of each All-in-one method in terms of the number of parameters (unit: million), number of floating-point operations (unit: GFLOPs) and average inference time per image (unit: milliseconds). The results show that the integrated network of the present invention has superior performance in terms of model scale, computational overhead and inference efficiency compared with the existing technology. Specifically, the network structure adopted by the present invention contains 1.03M learnable parameters, which is significantly less than other methods; the average inference time per image is 28.00 milliseconds, which is conducive to meeting the requirements of computational efficiency in practical applications. Although compared with the TransWeather method, the number of floating-point operations is slightly increased due to the integration of the forward physical propagation process, the overall computational overhead is still significantly lower than the real-time processing capability limit of the lensless microscopy imaging system, thereby ensuring the deployability and adaptability of the present invention in resource-constrained environments.

[0124] Table 4 Comparison of model complexity and inference efficiency of three all-in-one methods

[0125]

[0126] According to the technical solution of the present invention, this embodiment establishes a real lensless microscopic imaging system and collects diffraction patterns at different propagation distances to verify the applicability of the proposed algorithm in actual scenarios. In this imaging system, the central wavelength of the laser light source used is 532nm, and the imaging detector is a CMOS (model: DMK27AUJ003) device with a pixel pitch of 1.67μm and an imaging resolution of 1920×1080. The sample to be irradiated is derived from a stained tissue pathological section of thyroid papillary adenocarcinoma. The incident light is modulated by the aperture and irradiated to the sample, and the diffraction pattern formed by scattering is collected on the detector surface. The imaging resolution results in a difference in the sampling spacing in the x-axis and y-axis directions, that is, the fractional orders in the two directions are different, respectively denoted as p and y. x With p y This embodiment considers three specific examples, setting the distances between the light source and the sensor in the lensless imaging device to 1.716 mm, 3.912 mm, and 6.048 mm, respectively. The corresponding real observation patterns are collected under each distance condition. Subsequently, these patterns are input into the deep physical parameter integrated network proposed in this invention, and the phase recovery task is directly performed to reconstruct the corresponding sample images. The results are as follows:

[0127] 1. Diffraction propagation distance is 1.716 mm, such as Figure 7 As shown;

[0128] Corresponding fractional Fourier transform order: p x =0.3205,p y =0.1912

[0129] 2. Diffraction propagation distance is 3.912 mm, e.g. Figure 8 As shown;

[0130] Corresponding fractional Fourier transform order: p x =0.4055,p y =0.2510

[0131] 3. Diffraction propagation distance is 6.148 mm, e.g. Figure 9 As shown;

[0132] Corresponding fractional Fourier transform order: p x =0.5068,p y =0.3320

[0133] Verified by diffraction intensity patterns collected in an actual lensless imaging setup, the proposed deep physical parameter integrated network, thanks to its embedded forward propagation physics modeling and flexible parameter adaptation mechanism, effectively models and captures the complex variations in light waves during propagation, thereby improving reconstruction accuracy and detail restoration capabilities, enabling high-quality restoration using only a single observation image. The resulting reconstruction demonstrates significant advantages in image clarity, edge detail, and structural consistency, and is also applicable to lensless imaging applications involving multi-directional, heterogeneous fractional-order propagation conditions.

[0134] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the implementation methods of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. A phase retrieval method for a deep physical parameter integrated network for lensless microscopy, characterized in that: The following steps are involved: A physical forward model of lensless microscopy based on fractional Fourier transform is established to map the diffraction propagation distance of the light field into a controllable fractional order p. Constructing a progressive deep physical parameter integration network framework, the framework comprises M sequentially connected stages, each stage including: a physical model fidelity module and an image regularization update module; Physical model fidelity module: performs a forward fractional Fourier transform based on the fractional order p on the input image, replaces the amplitude with the actual collected observation intensity, and then performs an inverse fractional Fourier transform to output the updated estimated image; Image regularization update module: This module performs feature enhancement on the output of the physical model fidelity module, including: Adaptive proximal network: dynamically generates affine transformation parameters through fractional order p and noise intensity a to achieve conditional adaptive instance normalization; Adaptive compression-excitation network: The weights and biases of the fully connected layer are regulated by the fractional order p and noise intensity a, and channel attention weights are generated for feature weighting. A single-frame intensity observation image is input into the network, and the recovered complex amplitude information is directly output without the need for retraining for different imaging conditions.

2. The phase recovery method according to claim 1, wherein: The physical forward model satisfies: And p increases in the range [0,1]; Where d represents the diffraction propagation distance and λ represents the wavelength; And the intensity I recorded by the detector d The relationship with the sample complex light field U0 is: Among them, I d (x,y) represents the intensity recorded at the coordinate (x,y), F p represents the fractional Fourier transform operator, U0 represents the complex light field in the sample plane, and x, y represent the spatial coordinates of the sensor plane.

3. The phase recovery method according to claim 1, wherein: The adaptive proximal network achieves feature adaptation through the following formula: in, represents the conditional adaptive instance normalization operation of the jth layer in the kth stage, represents the input feature map, represents the channel mean of the feature map, represents the channel standard deviation of the feature map, represents the scaling factor generated dynamically by the fractional order p, represents the translation factor dynamically generated by the fractional order p; And the scaling factor and translation factor satisfy: in, represents the fully connected layer, and C represents the number of feature channels.

4. The phase recovery method according to claim 1, wherein: The adaptive compression-excitation network comprises: Compression operation: compress the feature map into a channel description vector through global average pooling; Excitation operation: It consists of two fully connected layers, whose weights and biases are dynamically generated by fractional order p regulation, and output channel weight vectors; Fusion operation: multiply the channel weights by the input features channel by channel, then perform convolution on the output and add a residual connection to retain the original information.

5. The phase recovery method according to claim 1, wherein: The noise intensity a is a Poisson noise parameter, and its value range is a∈[10,50]. The smaller the a value, the greater the noise intensity.

6. A phase retrieval device for lensless microscopic imaging, characterized in that: A method for implementing the phase recovery method according to any one of claims 1 to 5, comprising: A physical forward modeling unit configured to establish a light field propagation model based on fractional Fourier transform; Progressive processing unit: contains M stages in series, each stage includes: Physical model fidelity module: configured to perform fractional Fourier forward / inverse transform and intensity projection; Image regularization update module: This module includes an adaptive proximal network and an adaptive compression-excitation network, both of which dynamically control network parameters through fractional order p; Output unit: configured to recover the complex amplitude directly from a single frame of observation image.

7. A lensless microscopic imaging system, characterized in that: The phase recovery device according to claim 6 is integrated, wherein: The system uses a CMOS sensor to collect diffraction patterns, processes single-frame observation images through the device, and directly outputs complex amplitude reconstruction results of biological samples.

Images