Dark and weak extended target signal-to-noise ratio enhancement method based on time-space domain joint detection

Abstract

The present application belongs to the technical field of signal-to-noise ratio enhancement, and particularly relates to a dark and weak extended target signal-to-noise ratio enhancement method based on time-space domain joint detection. The method comprises the following steps: S1: performing preliminary noise reduction on a dark and weak extended target image sequence; S2: processing each dark and weak extended target image contained in the dark and weak extended target image sequence after preliminary noise reduction by using an adaptive weighted fusion noise reduction function; S3: inputting the dark and weak extended target image sequence after processing and deducting time domain noise into a trained deep convolutional neural network model for processing, to obtain a dark and weak extended target image sequence after deducting time-varying background noise; and S4: performing image enhancement on the dark and weak extended target image sequence after deducting time-varying background noise, to obtain a dark and weak extended target image sequence with high signal-to-noise ratio. The present application realizes effective suppression of time domain random noise, real-time estimation and elimination of sky light time-varying background, and greatly improves the detection capability of ground-based photoelectric detection equipment under strong sky light background in the daytime.

Description

Technical Field

[0001] This invention belongs to the field of signal-to-noise ratio enhancement technology, and particularly relates to a method for enhancing the signal-to-noise ratio of faint extended targets based on spatiotemporal joint detection. Background Technology

[0002] Currently, real-time, high-precision tracking, measurement, and identification of long-range, low-contrast, and dimly lit extended targets (such as aircraft, missiles, and drones) has wide-ranging applications. These targets are often in complex background interference environments, and due to factors such as long operating distance, weak target signals, and indistinct texture features, traditional detection methods struggle to achieve stable acquisition and continuous tracking.

[0003] The core challenges faced by traditional ground-based photoelectric detection systems during daytime observations are: strong background radiation from the sky can easily cause detector pixel saturation, greatly limiting the effective integration time of the detection system and reducing the accumulation efficiency of target photons; the large amount of photon shot noise introduced by the strong background can make it difficult for the system's signal-to-noise ratio to reach the detection threshold. Although traditional narrowband spectral filtering and spatial filtering methods can suppress background noise to some extent, they rely on the spectral differences between the target and the background and are prone to introducing structural noise that leads to target signal distortion, making it difficult to achieve high-fidelity extraction and daytime detection of long-distance, low-contrast extended targets. Due to these limitations, the daytime detection dynamic range of existing systems is typically below 60 dB, which cannot meet the all-day detection requirements for extremely low-contrast extended targets.

[0004] Among these methods, narrowband spectral filtering relies on prior information about the target spectrum; its picometer-level bandwidth, while suppressing skylight background, significantly attenuates the target signal energy. Polarization detection requires polarization modulators and multi-channel detectors, increasing system cost and complexity, and introducing calibration difficulties and reliability issues. Furthermore, time-varying polarization background noise from clouds, ground reflections, and airglow easily obscures the target's polarization characteristics, leading to a sharp decrease in detection probability. Shearing interferometry is extremely sensitive to optical path stability and environmental disturbances; micrometer-level vibrations or temperature drift can cause optical path changes that degrade or even fail the interference fringes. Therefore, it is difficult to apply to vibration platforms such as vehicle-mounted or airborne systems, and is typically limited to highly stable ground-based or space-based platforms. In summary, existing methods have significant limitations in improving the detection capabilities of telescope systems and cannot yet meet the application requirements for stable monitoring of faint space targets under strong daylight conditions. Summary of the Invention

[0005] In view of this, the present invention aims to provide a method for enhancing the signal-to-noise ratio of faint extended targets based on spatiotemporal joint detection, in order to solve the problem that the existing technology cannot meet the application requirements for stable monitoring of faint spatial targets under strong daylight background. The present invention achieves effective suppression of temporal random noise and real-time estimation and elimination of time-varying skylight background without losing the target signal energy, which greatly improves the detection capability of ground-based photoelectric detection equipment under strong daylight background.

[0006] To achieve the above objectives, the technical solution created by this invention is implemented as follows:

[0007] A signal-to-noise ratio enhancement method for faint extended targets based on joint spatiotemporal detection is applied to faint extended targets under strong daytime backgrounds, and specifically includes the following steps:

[0008] S1: Obtain a sequence of faint extended target images against a strong daylight background, and perform preliminary noise reduction on the faint extended target image sequence;

[0009] S2: An adaptive weighted fusion denoising function is used to process each of the weak extended target images contained in the initially denoised weak extended target image sequence to obtain a weak extended target image sequence after deducting temporal noise.

[0010] S3: After processing the dim extended target image sequence with time-domain noise removed, input it into the trained deep convolutional neural network model for further processing to obtain a dim extended target image sequence with time-varying background noise removed.

[0011] S4: Perform image enhancement on the dim extended target image sequence after removing time-varying background noise to obtain a dim extended target image sequence with high signal-to-noise ratio.

[0012] Furthermore, in step S1, spatial filtering is used to achieve preliminary noise reduction.

[0013] Furthermore, the faint expanded target images contained in the faint expanded target image sequence are arranged and numbered according to the acquisition time sequence.

[0014] Furthermore, step S2 specifically includes the following steps:

[0015] S21: Obtain the current overlay image that is superimposed with the faint expansion target image of frame m, where m is the maximum index in the current overlay image. Update the current overlay image by overlaying it with the faint expansion target image of frame (m+1). Use an adaptive weighted fusion denoising function to suppress the noise in the updated overlay image to obtain the overlay image after deducting temporal noise.

[0016] ;

[0017] in, For the updated overlay image, Let i and j be the pixel indices in the current faint expansion target image to be overlaid, W be the variance statistical feature, and g be the entropy statistical feature. Here, N represents the total number of rows in the currently overlaid weak and diffused target image, and P represents the total number of columns in the currently overlaid weak and diffused target image. As weight, The previous frame image of the weak, expanded target image to be overlaid. The next frame of the image to be overlaid is the dark, weak target image.

[0018] S22: Output the current superimposed image after removing temporal noise, replace the current superimposed image with the updated superimposed image and repeat step S21 until all faint expanded target images are superimposed to obtain a sequence of faint expanded target images after removing temporal noise.

[0019] Furthermore, in step S21, m starts from 1.

[0020] Furthermore, step S3 specifically includes the following steps:

[0021] S31: Extract the temporal difference signal of adjacent weakly extended target images and the spatial difference signal of adjacent pixels of each weakly extended target image from the weakly extended target image sequence after deducting temporal noise using the following formula:

[0022] ;

[0023] in, Let be the spatial difference signal in the width direction of the neighboring pixels of the faint extended target image at time t. For pixel position The faint extended target image at time t after deducting temporal noise. For pixel position The faint extended target image at time t after deducting temporal noise. For pixel position The faint expanded target image at time t after removing temporal noise. The temporal difference signal of adjacent dark and weak extended target images, Let h be the faint extended target image after removing temporal noise at time t-1, where h is the width direction of the image and v is the height direction of the image.

[0024] S32: Spatiotemporal joint multidimensional feature maps corresponding to each faintly extended target image in the sequence of faintly extended target images constructed and deducted from temporal noise based on spatial and temporal difference signals. :

[0025] ;

[0026] S33: Set a single-pixel threshold noise, and identify noisy pixels based on the relationship between the spatiotemporal joint multidimensional feature maps and the single-pixel threshold noise. :

[0027] ;

[0028] in, This is single-pixel threshold noise;

[0029] S34: Replace the noisy pixels in each spatiotemporal joint multidimensional feature map using the following formula to obtain the denoised spatiotemporal joint multidimensional feature map set:

[0030] ;

[0031] in, This is the denoised spatiotemporal joint multidimensional feature map at time t. To identify noisy pixels The median;

[0032] S35: Input the spatiotemporal joint multidimensional feature map set of noise-removed pixels into the trained deep convolutional neural network model for processing to obtain a sequence of dark and weak extended target images after deducting time-varying background noise.

[0033] Furthermore, in step S4, image enhancement includes adaptive histogram equalization, Laplacian pyramid enhancement, and gradient domain enhancement.

[0034] Furthermore, in step S3, the training set of the trained deep convolutional neural network model is the set of denoised spatiotemporal joint multidimensional feature maps obtained after processing the pre-collected dark extended target image sequence under strong daylight background in steps S1, S2 and S3. The loss function used to train the deep convolutional neural network model is the Huber loss function.

[0035] Furthermore, the deep convolutional neural network model employs a CNN network.

[0036] Compared with the prior art, the present invention can achieve the following beneficial effects:

[0037] (1) The present invention creates a method for enhancing the signal-to-noise ratio of weak extended targets based on joint temporal-spatial detection. It uses artificial intelligence decision theory to fuse multiple temporal noise features to achieve effective separation of the target and the background, thereby improving the signal-to-noise ratio of target detection.

[0038] (2) The method for enhancing the signal-to-noise ratio of faint extended targets based on spatiotemporal joint detection described in this invention achieves real-time and accurate noise reduction by training a deep learning convolutional neural network model end-to-end and predicting the background noise field from noisy observation data. Its core advantages are reflected in two aspects: first, it can preserve the target features in the original signal with high fidelity, including intensity, shape and texture; second, it has excellent robustness and generalization to various types of background undulations. Attached Figure Description

[0039] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0040] Figure 1 A flowchart illustrating the signal-to-noise ratio enhancement method for faint extended targets based on joint spatiotemporal detection, as described in an embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram illustrating the principle of the method for enhancing the signal-to-noise ratio of faint extended targets based on joint temporal-spatial detection, as described in an embodiment of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not constitute a limitation thereof.

[0043] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0044] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on this invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0045] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0046] The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0047] like Figures 1-2 As shown, this invention proposes a signal-to-noise ratio (SNR) enhancement method for faint extended targets based on spatiotemporal joint detection, applicable to faint extended targets against a strong daytime background. The method includes the following steps: S1: Acquire a sequence of faint extended target images against a strong daytime sky background and perform preliminary noise reduction on the sequence; S2: Process each faint extended target image in the pre-denoised sequence using an adaptive weighted fusion noise reduction function to obtain a sequence of faint extended target images with temporal noise removed; S3: Input the processed sequence of faint extended target images with temporal noise removed into a trained deep convolutional neural network model for further processing to obtain a sequence of faint extended target images with time-varying background noise removed; S4: Perform image enhancement on the sequence of faint extended target images with time-varying background noise removed to obtain a sequence of faint extended target images with a high SNR.

[0048] It should be noted that this invention relates to a spatiotemporal joint detection signal-to-noise ratio enhancement method for distant, dimly lit extended targets under conditions of strong daylight background flooding. This invention traces the physical sources of noise in photodetectors and establishes various noise impact mechanisms. It delves into the differences in statistical characteristics between noise signals and target signals in the time domain, constructing an adaptive (time-domain) weighted fusion noise reduction function. This aims to effectively suppress random noise while maximizing the preservation and enhancement of the true target signal, thereby improving the signal-to-noise ratio of the image sequence. Simultaneously, an artificial intelligence-based dynamic background prediction model is established to suppress interference from time-varying skylight backgrounds, achieving effective extraction of dimly lit targets.

[0049] The core objective of this invention is to address the critical challenge of low signal-to-noise ratio (SNR) and contrast when performing high-resolution imaging of faint, extended targets under strong background noise conditions (such as strong daylight). This invention is expected to enhance the detection capability and high-resolution imaging capability of ground-based photoelectric detection systems for long-distance, faint targets in strong daylight application scenarios such as solar adaptive optics, space telescope Earth observation, and airborne optical detection systems.

[0050] In some embodiments, in step S1, spatial filtering is used to achieve preliminary noise reduction.

[0051] It should be noted that spatial filtering includes median filtering.

[0052] In some embodiments, the faint expanded target images contained in the faint expanded target image sequence are arranged and numbered according to the acquisition time sequence.

[0053] It should be noted that the signal-to-noise ratio enhancement method for faint extended targets based on spatiotemporal joint detection in this invention is theoretically derived from the noise quantification data of various types provided by the offline noise testing system for photodetectors. The offline noise testing system for photodetectors consists of core components such as a high-precision uniform surface light source (integrating sphere light source), a blackbody radiation source, a spectrometer, an optical power meter, a precision optical adjustment frame, and attenuation plates. This system is mainly used to comprehensively evaluate various noise characteristics, dynamic range, and signal-to-noise ratio thresholds of photodetectors, providing crucial data support for subsequent system integration and the determination of practical application boundaries. Through standardized testing and multi-parameter analysis, a quantitative and multi-dimensional comprehensive evaluation of the noise performance of photodetectors is ultimately achieved. In terms of spatiotemporal joint correlation image processing, this invention mainly includes three parts: temporal random noise suppression, time-varying background noise suppression, and post-processing image enhancement.

[0054] The offline noise testing environment for photodetectors is equipped with a high-precision uniform radiation source, high and low temperature test chambers, and anechoic chamber facilities. This effectively suppresses stray light interference and controls temperature fluctuations, ensuring accurate and reliable noise test results. Furthermore, the offline noise testing system covers the visible light to long-wave infrared band, enabling precise measurement of key noise parameters (such as noise equivalent power NEP and detectivity D) across a wide spectrum and temperature range. It establishes noise influence mechanisms and simulates comprehensive noise characteristics under real-world application environments. This invention utilizes an established adaptive weighted fusion function and a trained deep convolutional neural network model to eliminate the influence of skylight background, thereby achieving the detection of weak targets in strong backgrounds.

[0055] In some embodiments, step S2 specifically includes the following steps:

[0056] S21: Obtain the current overlay image that is superimposed with the faint expansion target image of frame m, where m is the maximum index in the current overlay image. Update the current overlay image by overlaying it with the faint expansion target image of frame (m+1). Use an adaptive weighted fusion denoising function to suppress the noise in the updated overlay image to obtain the overlay image after deducting temporal noise.

[0057] ;

[0058] in, For the updated overlay image, Let i and j be the pixel indices in the current faint expansion target image to be overlaid, W be the variance statistical feature, and g be the entropy statistical feature. Here, N represents the total number of rows in the currently overlaid weak and diffused target image, and P represents the total number of columns in the currently overlaid weak and diffused target image. Weights (adjusted based on actual observation conditions). The previous frame image of the weak, expanded target image to be overlaid. The next frame of the image to be overlaid is the dark, weak target image.

[0059] Here, the current overlay image is overlaid and updated with the dim expansion target image of frame m+1. Therefore, the dim expansion target image to be overlaid is the dim expansion target image of frame m+1.

[0060] S22: Output the current superimposed image after removing temporal noise, replace the current superimposed image with the updated superimposed image and repeat step S21 until all faint expanded target images are superimposed to obtain a sequence of faint expanded target images after removing temporal noise.

[0061] In some embodiments, in step S21, m starts from 1.

[0062] It should be noted that by leveraging the repeatability of the target signal in the time domain and the statistical independence of noise, signal enhancement and noise suppression are achieved through the superposition of multiple frames. For independent noise points present in the image, effective suppression can be achieved through conventional multi-frame image superposition.

[0063] ;

[0064] If the noise exhibits temporal correlation and cross-correlation, causing its noise variance to decay at a rate lower than that of other noise domains, then the noise variance decays at a lower rate than that of other noise domains. This type of noise often constitutes a major component of an image, and conventional image overlay cannot effectively suppress it. Based on this problem, this invention proposes a digital integration algorithm for temporal multi-frame fusion.

[0065] By leveraging the statistical independence of temporal noise in long image sequences (such as the zero mean and temporal uncorrelation of Gaussian noise) and the temporal correlation of the target signal, an adaptive weighted fusion function is designed. By assigning higher weights to pixels with high temporal correlation, random noise is suppressed while preserving the true signal.

[0066] ;

[0067] Based on the quantitative relationship between detector noise and its operating parameters (such as integration time, gain, and temperature), the temporal feature statistics of pixels in long image sequences are analyzed (such as calculating temporal variance, covariance, or constructing autocorrelation functions) to quantify their temporal correlation, and this is used as the basis for weighted fusion. Based on the weighted fusion basis, the corresponding pixels in the image sequence are weighted averaged or optimally estimated to generate the final denoised image.

[0068] In some embodiments, step S3 specifically includes the following steps:

[0069] S31: Extract the temporal difference signal of adjacent weakly extended target images and the spatial difference signal of adjacent pixels of each weakly extended target image from the weakly extended target image sequence after deducting temporal noise using the following formula:

[0070] ;

[0071] in, Let be the spatial difference signal in the width direction of the neighboring pixels of the faint extended target image at time t. These are the pixel coordinates of the image. For pixel position The faint extended target image at time t after deducting temporal noise. For pixel position The faint extended target image at time t after deducting temporal noise. For pixel position The faint expanded target image at time t after removing temporal noise. The temporal difference signal of adjacent dark and weak extended target images, Let h be the faint extended target image after removing temporal noise at time t-1, where h is the width direction of the image and v is the height direction of the image.

[0072] S32: Spatiotemporal joint multidimensional feature maps corresponding to each faintly extended target image in the sequence of faintly extended target images constructed and deducted from temporal noise based on spatial and temporal difference signals. :

[0073] ;

[0074] S33: Set a single-pixel threshold noise, and identify noisy pixels based on the relationship between the spatiotemporal joint multidimensional feature maps and the single-pixel threshold noise. :

[0075] ;

[0076] in, This is single-pixel threshold noise;

[0077] S34: Replace the noisy pixels in each spatiotemporal joint multidimensional feature map using the following formula to obtain the denoised spatiotemporal joint multidimensional feature map set:

[0078] ;

[0079] in, This is the denoised spatiotemporal joint multidimensional feature map at time t. To identify noisy pixels The median;

[0080] S35: Input the spatiotemporal joint multidimensional feature map set of noise-removed pixels into the trained deep convolutional neural network model for processing to obtain a sequence of dark and weak extended target images after deducting time-varying background noise.

[0081] It should be noted that, in order to adapt to complex changes in the external environment (such as temperature changes, lighting changes, dynamic scene interference, etc.), deep convolutional neural networks (CNN) are used to extract the spatiotemporal features of image pixels, capture the background change patterns, model the time-varying background in the deep neural network model, predict the background of the current frame in real time, and compare the predicted background with the current image to remove outliers (noise) and retain the real signal, thereby achieving real-time suppression of background noise.

[0082] First, the difference signals between adjacent pixels are identified to achieve real-time suppression of time-varying background noise:

[0083] Background signals are continuous in the spatial domain, with gradual changes in the response values ​​of adjacent pixels. Noise signals (such as detector noise and sudden changes in illumination) appear as outliers in the spatial domain, with significant differences between adjacent pixels. Since noise exhibits high amplitude and low spatiotemporal correlation in the differential signal, a deep convolutional neural network can be established to calculate and analyze the spatiotemporal differential signals of adjacent pixels. By thresholding the amplitude of the differential signals, background and noise can be dynamically distinguished, potential noise can be effectively identified, and real-time suppression of time-varying noise can be achieved.

[0084] Traditional differential signal analysis methods based on manual thresholds have limited generalization ability in complex scenarios. Therefore, deep convolutional neural networks (CNNs) are introduced. Deep neural network models learn the spatial patterns of neighboring pixels through training, set various noise recognition thresholds, predict noise distribution, and further remove noise, thereby improving the robustness of time-varying background noise suppression models.

[0085] For a sequence of consecutively input multi-frame images, extract the spatial difference signal of adjacent pixels (e.g., horizontal, vertical, diagonal) in each frame and the temporal difference signal of adjacent frames:

[0086]

[0087] By combining spatial and temporal difference methods, a spatiotemporal joint multidimensional feature map is constructed:

[0088]

[0089] Predict the single-pixel noise threshold and establish the Huber loss function to constrain the difference between the noise recognition threshold and the true noise amplitude:

[0090]

[0091] in: The single-pixel noise threshold can be adaptively determined by the mean and standard deviation of the statistical differential signal.

[0092] Noise suppression: For pixels identified as noise, replace them with the neighborhood mean or median.

[0093]

[0094] Dynamic background compensation: Temporal filtering is applied to the noise-suppressed image to preserve the background trend.

[0095] In some embodiments, in step S4, image enhancement includes adaptive histogram equalization, Laplacian pyramid enhancement, and gradient domain enhancement.

[0096] It should be noted that Adaptive Histogram Equalization (CLAHE) divides the image to be processed into multiple N×N windows, performs histogram statistical distribution on the pixel values ​​in each window, and redistributes the pixel values. The original pixel values ​​are then mapped to new pixel values ​​through a cumulative distribution function, thereby enhancing the contrast of the target signal.

[0097] Laplacian pyramid enhancement: Gaussian pyramid decomposition is performed on the denoised image to divide the image into components of different frequencies. The high-frequency components (details) are enhanced and then reconstructed.

[0098] Gradient domain enhancement: Computing image gradients using the Sobel operator Introducing gradient amplification factor Implementing gradient boosting: By using the Poisson equation, the image is reconstructed from the enhanced gradient field to achieve image edge enhancement.

[0099] In some embodiments, in step S3, the training set of the trained deep convolutional neural network model is the set of spatiotemporal joint multidimensional feature maps of noise-removed pixels obtained after processing the pre-collected dark extended target image sequence under strong daylight background in steps S1, S2 and S3, and the loss function used to train the deep convolutional neural network model is the Huber loss function.

[0100] In some embodiments, the deep convolutional neural network model employs a CNN network.

[0101] The adaptive weighted fusion function and deep convolutional neural network model used in this invention are not limited to these and can be replaced by other neural network models with suitable functions.

[0102] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0103] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for enhancing the signal-to-noise ratio of faint extended targets based on joint spatiotemporal detection, applied to faint extended targets against a strong daytime background, characterized in that: Specifically, the steps include the following: S1: Obtain a sequence of faint extended target images against a strong daylight background, and perform preliminary noise reduction on the faint extended target image sequence; S2: An adaptive weighted fusion denoising function is used to process each of the weak extended target images contained in the initially denoised weak extended target image sequence to obtain a weak extended target image sequence after deducting temporal noise. S21: Obtain the current overlay image that is superimposed with the faint expansion target image of frame m, where m is the maximum index in the current overlay image. Update the current overlay image by overlaying it with the faint expansion target image of frame (m+1). Use an adaptive weighted fusion denoising function to suppress the noise in the updated overlay image to obtain the overlay image after deducting temporal noise. ； in, For the updated overlay image, Let i and j be the pixel indices in the current faint expansion target image to be overlaid, W be the variance statistical feature, and g be the entropy statistical feature. Here, N represents the total number of rows in the currently overlaid weak and diffused target image, and P represents the total number of columns in the currently overlaid weak and diffused target image. As weight, The previous frame image of the weak, expanded target image to be overlaid. The next frame of the currently overlaid, weakly extended target image; S22: Output the current superimposed image after removing temporal noise, replace the current superimposed image with the updated superimposed image and repeat step S21 until all dark expansion target images are superimposed to obtain a sequence of dark expansion target images after removing temporal noise. S3: After processing the dim extended target image sequence with time-domain noise removed, input it into the trained deep convolutional neural network model for further processing to obtain a dim extended target image sequence with time-varying background noise removed. S4: Perform image enhancement on the dim extended target image sequence after removing time-varying background noise to obtain a dim extended target image sequence with high signal-to-noise ratio.

2. The method for enhancing the signal-to-noise ratio of faint extended targets based on joint temporal-spatial detection according to claim 1, characterized in that: In step S1, spatial filtering is used to achieve preliminary noise reduction.

3. The method for enhancing the signal-to-noise ratio of faint extended targets based on joint temporal-spatial detection according to claim 1, characterized in that: The faint expanded target images in the sequence of acquired images are arranged and numbered according to the acquisition time.

4. The method for enhancing the signal-to-noise ratio of faint extended targets based on joint temporal-spatial detection according to claim 1, characterized in that: In step S21, m starts from 1.

5. The method for enhancing the signal-to-noise ratio of faint extended targets based on joint temporal-spatial detection according to claim 1, characterized in that: Step S3 specifically includes the following steps: S31: Extract the temporal difference signal of adjacent weakly extended target images and the spatial difference signal of adjacent pixels of each weakly extended target image from the weakly extended target image sequence after deducting temporal noise using the following formula: ； in, Let be the spatial difference signal in the width direction of the neighboring pixels of the faint extended target image at time t. These are the pixel coordinates of the image. For pixel position The faint extended target image at time t after deducting temporal noise. For pixel position The faint extended target image at time t after deducting temporal noise. For pixel position The faint expanded target image at time t after removing temporal noise. The temporal difference signal of adjacent dark and weak extended target images, Let h be the faint extended target image after removing temporal noise at time t-1, where h is the width direction of the image and v is the height direction of the image. S32: Spatiotemporal joint multidimensional feature maps corresponding to each faintly extended target image in the sequence of faintly extended target images constructed and deducted from temporal noise based on spatial and temporal difference signals. : ； S33: Set a single-pixel threshold noise, and identify noisy pixels based on the relationship between the spatiotemporal joint multidimensional feature maps and the single-pixel threshold noise. : ； in, This is single-pixel threshold noise; S34: Replace the noisy pixels in each spatiotemporal joint multidimensional feature map using the following formula to obtain the denoised spatiotemporal joint multidimensional feature map set: ； in, This is the denoised spatiotemporal joint multidimensional feature map at time t. To identify noisy pixels The median; S35: Input the spatiotemporal joint multidimensional feature map set of noise-removed pixels into the trained deep convolutional neural network model for processing to obtain a sequence of dark and weak extended target images after deducting time-varying background noise.

6. The method for enhancing the signal-to-noise ratio of faint extended targets based on joint temporal-spatial detection according to claim 1, characterized in that: In step S4, image enhancement includes adaptive histogram equalization, Laplacian pyramid enhancement, and gradient domain enhancement.

7. The method for enhancing the signal-to-noise ratio of faint extended targets based on joint temporal-spatial detection according to claim 1, characterized in that: In step S3, the training set of the trained deep convolutional neural network model is the set of denoised spatiotemporal joint multidimensional feature maps obtained after processing the pre-collected image sequence of dark extended targets under strong daylight background in steps S1, S2 and S3. The loss function used to train the deep convolutional neural network model is the Huber loss function.

8. The method for enhancing the signal-to-noise ratio of faint extended targets based on joint temporal-spatial detection according to claim 7, characterized in that: The deep convolutional neural network model uses a CNN network.

Images