Snapshot compressive spatio-temporal encoding i-coach four-dimensional imaging system and method based on time-varying phase mask

Abstract

The application belongs to the technical field of optics and artificial intelligence, and particularly relates to a snapshot type compressed space-time coding I-COACH four-dimensional imaging system and method based on a time-varying phase mask. The single exposure time compression mechanism of SCI is introduced into the I-COACH imaging framework, a group of mutually different and depth-related PSF libraries are constructed by a time-varying coding phase mask. The object information of different time frames and depth planes is respectively coded by the corresponding PSF and superimposed into a compressed I-COACH hologram in a single integral exposure of the detector. At the reconstruction end, the space-time coding imaging process of I-COACH is explicitly expressed as a forward physical model, and is introduced as a constraint into the reconstruction network to separate multiple time channels and depth channels from a single measurement. Through 10-frame dynamic scene reconstruction experiments, PSF correlation analysis and reconstruction algorithm comparison, the time coding compression and reconstruction imaging capability of the method are verified. Further, multi-depth, multi-time numerical simulation evaluates the feasibility of the method in incoherent holographic four-dimensional imaging.

Description

Technical Field

[0001] This invention belongs to the fields of optics and artificial intelligence, specifically relating to a snapshot-based compressed spatiotemporal coding I-COACH four-dimensional imaging system and method based on time-varying phase masks. Background Technology

[0002] Interferenceless Coded Aperture Correlation Holography (I-COACH) is a non-interferometric, single-exposure 3D imaging technique. This method encodes the 3D information of spatially incoherent objects into 2D intensity holograms and recovers object information at different depths by cross-correlation with a pre-calibrated point spread function (PSF). Since it eliminates the need for a beam splitter to create dual-wavelength self-interference, I-COACH features a simple optical path and high stability; its depth-correlated PSF response provides axial resolution. However, traditional I-COACH records only a full-frame hologram at a single moment per exposure, thus limiting the temporal resolution of dynamic imaging to camera frame rate and readout speed. For rapidly changing scenes, temporal undersampling not only causes motion blur but also significantly increases data transmission and storage burden due to continuous acquisition of full-frame holograms.

[0003] In existing technologies, increasing the camera frame rate can alleviate the problem of insufficient temporal resolution, but this usually comes at the cost of increased hardware costs and data throughput. Video frame interpolation estimates intermediate states from adjacent frames, and the newly added frames are not directly obtained measurement information during the exposure process. Therefore, ghosting, blurring, or pseudo-structures may occur during rapid movement, occlusion, or abrupt changes in detail. Snapshot Compressive Imaging (SCI) modulates multiple temporal frames within a single exposure using a rapidly changing random pixel sampling coded mask, superimposes them into a single compressed measurement, and then combines it with relevant reconstruction algorithms to recover the high-speed sequence. This front-end physical encoding can bypass the limitation of high-speed frame-by-frame readout by the camera and reduce data transmission and storage overhead. However, existing SCI uses an intensity-coded mask to randomly sample each frame of the image, so it mainly focuses on xyt information and usually does not have the inherent axial resolution capability of I-COACH, making it difficult to directly obtain xyzt information of dynamic 3D scenes. To address this, this invention provides a snapshot-based compressed spatiotemporal coding I-COACH four-dimensional imaging system and method based on a time-varying phase mask. Summary of the Invention

[0004] The purpose of this invention is to provide a snapshot-based compressed spatiotemporal coding I-COACH four-dimensional imaging system and method based on time-varying phase masks. This invention introduces the single-exposure time compression mechanism of SCI into the I-COACH imaging framework, proposing a snapshot-based compressed spatiotemporal coding I-COACH imaging method that effectively overcomes the limitation of traditional I-COACH, which can only encode a single time frame of the 3D scene within a single exposure. This imaging method no longer uses intensity masks to randomly sample pixels from 2D images of the scene at different times. Instead, it constructs a set of mutually different and depth-related PSF libraries using time-varying coded phase masks. Object information in different time frames and depth planes is encoded by the corresponding PSFs and superimposed into a compressed I-COACH hologram within a single integral exposure of the detector. Therefore, a single measurement image contains information from multiple time frames and multiple depth-coded images, thus preserving the 3D imaging capability of I-COACH. At the reconstruction end, this invention explicitly describes the spatiotemporal coding imaging process of I-COACH as a forward physical model and introduces it as a constraint into the reconstruction network, separating multiple time channels and depth channels from a single measurement. The proposed method's temporal coding compression and reconstruction imaging capabilities were verified through a 10-frame dynamic scene reconstruction experiment, PSF correlation analysis, and reconstruction algorithm comparison. Further multi-depth, multi-time numerical simulations were used to evaluate its feasibility in incoherent holographic four-dimensional imaging.

[0005] The specific technical solution adopted by this invention is as follows: The snapshot-based compressed spatiotemporal coding I-COACH four-dimensional imaging system and method based on time-varying phase masks is characterized by: using a light-emitting diode (LED; Thorlabs M625L4-C1, power 630mW) with a center wavelength of 625nm as a spatial incoherent light source; lenses L1 and L2 together with the aperture to form a collimation and beam-shrinking system for illuminating a digital micromirror device (DMD; Fldiscovery F4300, resolution 1920×1080 pixels, pixel size 10.8μm); the digital micromirror device displays dynamically changing object patterns in time sequence to simulate high-speed motion scenes; subsequently, the object light passes sequentially through lens L3 and a polarizer P with a transmission axis of 0°. After passing through the unpolarized beam splitter (BS), the light wave reaches the pure phase spatial light modulator (SLM; Holoeye PLUTO, resolution 1920×1080 pixels, pixel size 8μm, refresh rate 60Hz); the compressed coded image is captured by the camera (LUCIDTRI050S-MC, resolution 4096×3000 pixels, pixel size 3.45μm, full-frame frame rate 9Hz), and the camera exposure time is set to 200ms; the phase spatial light modulator and the camera are synchronized by rising edge triggering, so that the camera completes a single exposure during the continuous modulation of the phase spatial light modulator, and compresses and records the coded scene at multiple moments into a single measurement image.

[0006] The snapshot-based compressed spatiotemporal coding I-COACH four-dimensional imaging method based on time-varying phase masks performs the following steps sequentially after the four-dimensional imaging system acquires the measurement image: Step 1: Time-varying phase encoding and PSF calibration; In the phase mask design phase, a set of target PSF intensity distributions with different shape characteristics are first defined to enhance the distinguishability between different time channels. The GS algorithm is then used to generate the phase mask.

[0007] Specifically, using a random phase as the initial input, the amplitude portion of the current complex amplitude is replaced with the target PSF amplitude in the frequency domain, while preserving its phase information. Then, an inverse Fourier transform is applied back to the spatial domain, and the phase distribution is updated. After multiple iterations, a phase mask capable of generating a PSF of a specific shape in the detector plane can be obtained.

[0008] In step 1, before formally acquiring dynamic scenes, point targets are used to calibrate the system. The positions of the light source, imaging optical path, spatial light modulator, camera, and system parameters are kept consistent with the actual acquisition process. Encoded phase masks corresponding to each moment are sequentially loaded onto the spatial light modulator, and the camera records the corresponding point spread functions.

[0009] To improve the separability between different time channels, the coded phase mask can be screened or optimized based on the cross-correlation between the spread functions at each point, so that the spread functions at different points have low cross-correlation, thereby reducing inter-frame crosstalk in the subsequent reconstruction process. This invention uses the normalized cross-correlation coefficient shown below. ρ Characterizing the correlation between any two PSFs: ; In the formula, Indicates the first i PSF at pixel coordinates Strength at that location, Indicates the first j PSF at pixel coordinates Strength at that location, and The first i The and the first j Average pixel intensity of each PSF. ρ The larger the value, the more similar the two PSFs. For those containing... N frame( N The encoded sequence of ≥2) will be further... N Between each pair of PSF ρ Taking the average, the average correlation coefficient of this PSF set is defined as: ; In the formula, N Indicates the number of encoded PSFs. The larger the value, the higher the overall correlation of the PSF group, the worse the channel separability, and the easier it is to generate crosstalk during reconstruction.

[0010] Step 2: Single-exposure compressed acquisition: Within one exposure cycle, multiple time frames of the dynamic target are sequentially encoded by the corresponding PSF, and the sensor ultimately records only one compressed observation image.

[0011] In step 2, the forward model can be represented as: in, Y This represents the compression measurement obtained from a single exposure. O k For the first k Object intensity distribution over time frames, PSF k This is the point spread function for the corresponding time channel. Represents a two-dimensional spatial convolution operation. This represents the model error and noise term.

[0012] Step 3: Physics-Driven Reconstruction: Construct a physics-driven reconstruction network, taking the compressed measurement value Y and the PSF at each time step as input. Explicitly introduce a forward physics model, data consistency constraints, and temporal prior constraints into the network. Through iterative expansion or hybrid optimization, reconstruct the dynamic sequence of multiple time frames, thereby restoring the target's... xyzt Four-dimensional information.

[0013] In step 3, the network training process employs a joint optimization of reconstruction loss, data consistency loss, structural similarity loss, and temporal smoothing loss.

[0014] Step 4: Experimental Testing; The reconstruction performance of this invention is verified through dynamic target testing, PSF correlation testing, and comparison with traditional iterative algorithms such as TWIST, GAP-TV, and ADMM. According to the experimental results, in a 10-frame dynamic reconstruction task, the method of this invention can achieve an average PSNR of 29.876dB and an average SSIM of 0.851, which is better than the comparison algorithms, indicating that this method has good applicability in single-exposure dynamic four-dimensional imaging scenarios.

[0015] The technical effects achieved by this invention are as follows: This invention uses time-varying phase encoding and single-exposure compression acquisition mechanism to encode the dynamic information of multiple time frames into a single compressed measurement value, avoiding the decrease in temporal resolution caused by multiple exposures, and is suitable for fast dynamic scenes.

[0016] This invention reduces crosstalk between time channels by constraining and optimizing the correlation of PSFs in different time frames, enabling clearer separation of information from each frame in compressed observation and improving reconstruction stability.

[0017] This invention introduces a physical forward model into the reconstruction network, which enhances interpretability and generalization ability while retaining the efficient expressive power of deep learning. Compared with pure data-driven methods, it is more suitable for dealing with different targets, different motion speeds and different noise conditions.

[0018] According to the experimental results, the present invention achieved high average PSNR and SSIM in the 10-frame dynamic reconstruction task, and has significant advantages over baseline methods such as TWIST, GAP-TV, and ADMM, indicating that the scheme has practical application value in single-exposure dynamic four-dimensional (z,y,z,t) imaging. Attached Figure Description

[0019] Figure 1 The overall flowchart for I-COACH, a single exposure time compression encoding method.

[0020] Figure 2 This is a diagram of a four-dimensional imaging system.

[0021] Figure 3 This is a schematic diagram of the compressed recording and reconstruction of a dynamic scene of 10 frames under a single exposure.

[0022] Figure 4 This diagram illustrates the impact of encoding PSF correlation on dynamic reconstruction performance.

[0023] Figure 5 This diagram illustrates the comparison of results from different reconstruction methods.

[0024] Figure 6 This is a schematic diagram of 4D reconstruction of a multi-depth dynamic scene. Detailed Implementation

[0025] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0026] like Figure 1 As shown, the snapshot-based compressed spatiotemporal coding I-COACH four-dimensional imaging system based on a time-varying phase mask uses a light-emitting diode with a center wavelength of 625nm as a spatially incoherent light source. Lenses L1 and L2, together with the aperture, form a collimation and beam-shrinking system to illuminate a digital micromirror device (DMM). The DMM displays dynamically changing object patterns in chronological order to simulate high-speed motion scenes. Subsequently, the object light passes sequentially through lens L3 and a polarizer P with a transmission axis of 0°. After passing through the non-polarizing beam splitter BS, the light wave reaches the pure phase spatial light modulator. The compressed coded image is acquired by a camera with an exposure time set to 200ms. Synchronization between the phase spatial light modulator and the camera is achieved using a rising edge triggering method, allowing the camera to complete a single exposure during the continuous modulation of the phase spatial light modulator and compressing and recording the coded scene at multiple moments into a single measurement image.

[0027] In this embodiment, the LED model is Thorlabs M625L4-C1, and the power is 630mW; The digital micromirror device (DMD) is model FldiscoveryF4300, with a resolution of 1920×1080 pixels and a pixel size of 10.8μm; The pure phase spatial light modulator (SLM) is model HoloeyePLUTO, with a resolution of 1920×1080 pixels, a pixel size of 8μm, and a refresh rate of 60Hz. The camera model is LUCIDTRI050S-MC, with a resolution of 4096×3000 pixels, a pixel size of 3.45μm, and a full-frame frame rate of 9Hz. The snapshot-based compressed spatiotemporal coding I-COACH four-dimensional imaging method based on time-varying phase masks performs the following steps sequentially after the four-dimensional imaging system acquires the measurement image: Step 1: Time-varying phase encoding and PSF calibration; In the phase mask design phase, a set of target PSF intensity distributions with different shape characteristics are first defined to enhance the distinguishability between different time channels. The GS algorithm is then used to generate the phase mask.

[0028] Specifically, using a random phase as the initial input, the amplitude portion of the current complex amplitude is replaced with the target PSF amplitude in the frequency domain, while preserving its phase information. Then, an inverse Fourier transform is applied back to the spatial domain, and the phase distribution is updated. After multiple iterations, a phase mask capable of generating a PSF of a specific shape in the detector plane can be obtained.

[0029] In step 1, before formally acquiring dynamic scenes, point targets are used to calibrate the system. The positions of the light source, imaging optical path, spatial light modulator, camera, and system parameters are kept consistent with the actual acquisition process. Encoded phase masks corresponding to each moment are sequentially loaded onto the spatial light modulator, and the camera records the corresponding point spread functions.

[0030] To improve the separability between different time channels, the coded phase mask can be screened or optimized based on the cross-correlation between the spread functions at each point, so that the spread functions at different points have low cross-correlation, thereby reducing inter-frame crosstalk in the subsequent reconstruction process. This invention uses the normalized cross-correlation coefficient shown below. ρ Characterizing the correlation between any two PSFs: ; In the formula, Indicates the first i PSF at pixel coordinates Strength at that location, Indicates the first j PSF at pixel coordinates Strength at that location, and The first i The and the first j Average pixel intensity of each PSF. ρ The larger the value, the more similar the two PSFs. For those containing... N frame( N The encoded sequence of ≥2) will be further... N Between each pair of PSF ρ Taking the average, the average correlation coefficient of this PSF set is defined as: ; In the formula,N Indicates the number of encoded PSFs. The larger the value, the higher the overall correlation of the PSF group, the worse the channel separability, and the easier it is to generate crosstalk during reconstruction.

[0031] Step 2: Single-exposure compressed acquisition: Within one exposure cycle, multiple time frames of the dynamic target are sequentially encoded by the corresponding PSF, and the sensor ultimately records only one compressed measurement image.

[0032] In step 2, the forward model can be represented as: in, Y This represents the compression measurement obtained from a single exposure. O k For the first k Object intensity distribution over time frames, PSF k This is the point spread function for the corresponding time channel. Represents a two-dimensional spatial convolution operation. This represents the model error and noise term.

[0033] Step 3: Physics-Driven Reconstruction: Construct a physics-driven reconstruction network, taking the compressed measurement value Y and the PSF at each time step as input. Explicitly introduce a forward physics model, data consistency constraints, and temporal prior constraints into the network. Through iterative expansion or hybrid optimization, reconstruct the dynamic sequence of multiple time frames, thereby restoring the target's... xyzt Four-dimensional information.

[0034] In step 3, the network training process employs a joint optimization of reconstruction loss, data consistency loss, structural similarity loss, and temporal smoothing loss.

[0035] Step 4: Experimental Testing; The reconstruction performance of this invention is verified through dynamic target testing, PSF correlation testing, and comparison with traditional iterative algorithms such as TWIST, GAP-TV, and ADMM. According to the experimental results, in a 10-frame dynamic reconstruction task, the method of this invention can achieve an average PSNR of 29.876dB and an average SSIM of 0.851, which is better than the comparison algorithms, indicating that this method has good applicability in single-exposure dynamic four-dimensional imaging scenarios.

[0036] Figure 1 The overall workflow for single-exposure time compression coding (I-COACH) is as follows: (a) single-exposure time compression coding; (b) coding the phase mask and its corresponding PSF; (c) dynamic sequence reconstruction network based on the forward physics model.

[0037] Figure 2This is a diagram of a four-dimensional imaging system. L: Lens; DMD: Digital Micromirror Device; BS: Beam Splitter; P: Polarizer; SLM: Spatial Light Modulator.

[0038] Figure 3 Compressed recording and reconstruction of a 10-frame dynamic scene under a single exposure. (a) Compressed measurement obtained from a single exposure; (b) System PSF corresponding to 10 time channels; (c) 10-frame dynamic sequence reconstructed from the single-image compressed measurement; Figure 4 The impact of coded PSF correlation on dynamic reconstruction performance. (a) Reconstruction results of the third frame corresponding to the four PSFs; (b) Correlation matrix and average correlation coefficient of each PSF; (c) Relationship between average correlation coefficient and PSNR and SSIM; Figure 5 Comparison of results from different reconstruction methods. From left to right: compressed measurement, true value, TWIST, GAP-TV, ADMM, and the result of the method presented in this paper;

[0039] Figure 6 4D reconstruction of multi-depth dynamic scenes. (a) Compression encoding and reconstruction at four time points in three depth planes; (b) Corresponding four-dimensional rendering results; Combination Figure 1 and Figure 2 As can be seen, this invention unifies time-varying phase masks, I-COACH 3D encoded imaging, single-exposure compressed acquisition, and physically driven reconstruction into a single imaging chain. The phase spatial light modulator continuously loads different phase masks during a single camera exposure, enabling the spatial depth information corresponding to different time frames to be encoded using different point spread functions, and finally superimposed to form a compressed measurement image. Therefore, the system does not need to expose and record holograms for each time frame individually, significantly reducing the time overhead, data volume, and storage pressure caused by multiple acquisitions while maintaining the I-COACH axial imaging capability, making it suitable for transient recording in rapidly changing scenes.

[0040] Combination Figure 3 As can be seen, this invention can recover a dynamic sequence of 10 time frames from a compressed record obtained from a single exposure. The reconstruction results can respectively present the structural changes of the target at times T1 to T10, indicating that this invention does not only enhance a single static image, but can decode and separate multiple time channels to recover the dynamic evolution process of the target. Compared with the traditional dynamic I-COACH scheme that requires continuous acquisition of multiple holograms, this invention can reduce inter-frame misalignment and motion blur caused by the displacement of moving targets between adjacent exposures, thus making it more suitable for imaging targets with high speed, deformation, or rapid changes.

[0041] Combination Figure 4 It can be seen that the correlation between PSFs corresponding to different time channels is an important factor affecting reconstruction quality and inter-frame crosstalk. When the average correlation coefficient of different PSFs is low, the coding features of each time channel are easier to distinguish, and the crosstalk and artifacts in the reconstruction result are correspondingly reduced; when the PSF correlation increases, the information of different time frames is more likely to overlap, and the reconstruction quality deteriorates. Therefore, this invention transforms the separability of time channels into measurable and optimizable design parameters by screening and optimizing the coded phase mask and using the average correlation coefficient of PSFs as a quantitative evaluation index. The average correlation coefficient of the PSF set examined in the experiment covers approximately 0.136 to 0.444. Figure 4 The results shown verify that reducing PSF correlation can improve dynamic reconstruction performance, thus providing a clear basis for the design and selection of phase masks.

[0042] Combination Figure 5 As shown in Table 1, the average PSNR obtained by the physical-driven reconstruction network in this invention is 29.876 dB, and the average SSIM is 0.851. This represents a significant improvement in reconstruction quality compared to traditional iterative algorithms such as TWIST, GAP-TV, and ADMM. This result demonstrates that the reconstruction algorithm proposed in this invention not only improves the overall signal-to-noise quality of the reconstructed image but also better preserves the contours, edges, and internal structure of the target, reducing structural defects, background artifacts, and inter-frame crosstalk that are prone to occur in traditional iterative algorithms under strong compression conditions. Figure 5 Visual comparisons further illustrate that, under the same single-exposure compressed observation and the same encoded PSF, the reconstruction algorithm proposed in this invention can more completely recover the position and shape of sparse targets, with clearer target structures and cleaner background areas. This is because this invention explicitly introduces the forward imaging model, calibrated PSF, and data consistency constraints into the reconstruction process, making the network output subject to both optical propagation laws and observation data constraints. Therefore, even with a high compression rate in the observation data, the reconstruction results still exhibit good physical consistency and structural fidelity.

[0043] Combination Figure 6 As can be seen, this invention can jointly reconstruct dynamic targets located in different depth planes, simultaneously providing the target's lateral position, axial depth, and time-varying motion state at different time points. Therefore, the information output by the system is no longer limited to two-dimensional video sequences, but achieves the recovery of four-dimensional dynamic information (xyzt) from single-frame compressed measurements. This capability makes this invention applicable to moving targets with significant depth variations and dynamic three-dimensional scenes requiring simultaneous analysis of spatial position and temporal evolution.

[0044] The number of reconstruction frames N in this invention can be adjusted according to the target speed, sensor frame rate, and coding complexity, and can be 4, 6, 8, 10, or more frames. As N changes, the correlation constraints of the PSF, the network input size, and the loss weights can be adjusted synchronously to adapt to different application scenarios.

[0045] The reconstruction network is not limited to a fixed structure. It can be replaced with an expanded network, U-Net structure, Transformer structure or a hybrid structure, depending on the implementation needs, as long as it can complete the mapping from compressed observations to dynamic four-dimensional results under the constraints of the physical model.

[0046] In summary, this invention simultaneously achieves single-exposure capability, high temporal resolution, 3D depth acquisition, 4D dynamic reconstruction, low inter-frame crosstalk, and high reconstruction quality, demonstrating clear technological advancement and practical application value. The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A snapshot-based compressed spatiotemporal coding I-COACH four-dimensional imaging system based on time-varying phase masks, characterized in that: A light-emitting diode with a center wavelength of 625nm is used as a spatial incoherent light source; lenses L1 and L2, together with the aperture, form a collimation and beam-shrinking system to illuminate the digital micromirror device (DMM); the DMM displays dynamically changing object patterns in chronological order to simulate high-speed motion scenes; subsequently, the object light passes sequentially through lens L3 and a polarizer P with a transmission axis of 0°; after passing through the non-polarizing beam splitter, the light wave reaches the pure phase spatial light modulator; the compressed coded image is acquired by the camera, with the camera exposure time set to 200ms; the phase spatial light modulator and the camera are synchronized using a rising edge triggering method, enabling the camera to complete a single exposure during the continuous modulation of the phase spatial light modulator, and compressing and recording the coded scene at multiple moments into a single measurement image.

2. A snapshot-based compressed spatiotemporal coding I-COACH four-dimensional imaging method based on time-varying phase masks, characterized in that: After acquiring the measurement image using the four-dimensional imaging system of claim 1, the following steps are performed sequentially: Step 1: Time-varying phase encoding and PSF calibration; In the phase mask design stage, a set of target PSF intensity distributions with different shape characteristics are first set to enhance the distinguishability between different time channels; the GS algorithm is used to generate the phase mask. Specifically, using a random phase as the initial input, the amplitude portion of the current complex amplitude is replaced with the target PSF amplitude in the frequency domain while retaining its phase information; then, the phase distribution is updated by returning to the spatial domain through an inverse Fourier transform; after multiple iterations, a phase mask that can generate a PSF of a specific shape in the detector plane can be obtained. Step 2: Single-exposure compressed acquisition: Within one exposure cycle, multiple time frames of the dynamic target are sequentially encoded by the corresponding PSF, and the sensor ultimately records only one compressed observation image; Step 3: Physics-Driven Reconstruction: Construct a physics-driven reconstruction network, taking the compressed measurement value Y and the PSF at each time step as input. Explicitly introduce a forward physics model, data consistency constraints, and temporal prior constraints into the network. Through iterative expansion or hybrid optimization, reconstruct the dynamic sequence of multiple time frames, thereby restoring the target's... xyzt Four-dimensional information.

3. The snapshot-based compressed spatiotemporal coding I-COACH four-dimensional imaging method based on time-varying phase mask according to claim 2, characterized in that: In step 1, before formally acquiring dynamic scenes, point targets are used to calibrate the system; the positions of the light source, imaging optical path, spatial light modulator, camera position and system parameters are kept consistent with the actual acquisition process, and the corresponding encoded phase masks at each time moment are loaded onto the spatial light modulator in sequence, and the corresponding point spread functions are recorded by the camera respectively. To improve the separability between different time channels, the coded phase mask can be screened or optimized based on the cross-correlation between the spread functions at each point, so that the spread functions at different points have low cross-correlation, thereby reducing inter-frame crosstalk in the subsequent reconstruction process; this invention uses the normalized cross-correlation coefficient shown below. ρ Characterizing the correlation between any two PSFs: ； In the formula, Indicates the first i PSF at pixel coordinates Strength at that location, Indicates the first j PSF at pixel coordinates Strength at that location, and The first i The and the first j Average pixel intensity of each PSF; ρ The larger the value, the more similar the two PSFs; for those containing N frame( N The encoded sequence of ≥2) will be further... N Between each pair of PSF ρ Taking the average, the average correlation coefficient of this PSF set is defined as: ； In the formula, N Indicates the number of encoded PSFs. The larger the value, the higher the overall correlation of the PSF group, the worse the channel separability, and the easier it is to generate crosstalk during reconstruction.

4. The snapshot-based compressed spatiotemporal coding I-COACH four-dimensional imaging method based on time-varying phase mask according to claim 2, characterized in that: In step 2, the forward model can be represented as: in, Y This represents the compression measurement obtained from a single exposure. O k For the first k Object intensity distribution over time frames, PSF k This is the point spread function for the corresponding time channel. Represents a two-dimensional spatial convolution operation. This represents the model error and noise term.

5. The snapshot-based compressed spatiotemporal coding I-COACH four-dimensional imaging method based on time-varying phase mask according to claim 2, characterized in that: In step 3, the network training process employs a joint optimization of reconstruction loss, data consistency loss, structural similarity loss, and temporal smoothing loss.

Images