178 results about "Speckle noise" patented technology

The invention relates to the technical field of optical lens defect detection, and particularly discloses an optical lens laser-induced damage on-line detection method and system, a linear polarizer and a narrow-band filter are connected in series in a detection laser light path, non-target polarized light is filtered through polarization direction matching, meanwhile, interference of the environment and scattered light is inhibited through the narrow-band filter, and the laser-induced damage on-line detection system is obtained. According to the method, external source noise such as ambient light and non-target polarized light is reduced through polarization matching and narrow-band filtering, randomness of speckle noise is offset through multi-angle collection and signal fusion in a visual detection mode, exposure is dynamically adjusted, an image is fused to improve the signal-to-noise ratio, noise and real signals are decoupled through self-supervised learning, and the real-time performance of the system is improved. Dust interference is removed in combination with morphological operation; a deep learning reasoning model enhances the damage identification capability in a noise environment, defocusing blur caused by mechanical vibration is eliminated through phase conjugate correction, and the influence of photoelectric conversion noise, dust interference and the like on the detection efficiency and accuracy is remarkably reduced through multi-link linkage.

The invention relates to the technical field of unmanned aerial vehicle control, in particular to an unmanned aerial vehicle dynamic obstacle avoidance method based on multi-sensor fusion. Comprising the steps of receiving an image sequence, a depth point cloud set, a pose state parameter and a flight speed parameter; calculating an illumination distortion gradient value and a speckle noise entropy value based on the image sequence and the depth point cloud set; performing coordinate registration and feature extraction on the image sequence, the depth point cloud set and the pose state parameter, and outputting a visual feature set and a radar feature set; generating a weighted visual feature set and a weighted radar feature set through a weight adjustment function, and generating a joint heterogeneous feature tensor through alignment splicing; and generating a body attitude and thrust control instruction by using the combined heterogeneous feature tensor through a probability prediction model and a model prediction control algorithm device. According to the method, through cross-domain multiplexing and deep coupling of the body physical parameters in the data stream, error accumulation and decision delay caused by multi-stage series calculation are avoided, and global performability of an obstacle avoidance task is facilitated.

The invention relates to the technical field of radar remote sensing image processing, in particular to a multi-scale synthetic aperture radar flood detection method and device, and the method comprises the steps: collecting a plurality of flood disaster SAR images of a flood region, and carrying out the preprocessing of the images, so as to obtain a standard flood disaster SAR image; dividing standard flood disaster SAR images, and constructing training, verification and test data sets; based on a multi-scale feature extraction network and a multi-head self-attention mechanism, constructing a multi-scale SAR flood detection network model, training the multi-scale SAR flood detection network model by using the training data set and the verification data set, and inputting the test data set into the trained multi-scale SAR flood detection network model, therefore, the influence of speckle noise is effectively suppressed, the capability of distinguishing flood from confusion-prone ground features in a complex scene is improved, and the accuracy and robustness of SAR image flood detection are improved.

The invention belongs to the technical field of image enhancement and recognition, and particularly relates to a radar image intelligent enhancement and recognition method and system based on multi-model fusion, and the method comprises the steps: carrying out the adaptive suppression of speckle noise of an original radar image; feature point detection is carried out, robust transformation matrix estimation and adaptive contrast enhancement are carried out, and a corrected and enhanced image is output; utilizing the generative adversarial network and multi-loss function collaborative constraint to obtain a texture reconstruction image; establishing an image-semantic double-flow network architecture, performing cross-modal attention fusion to obtain a fusion feature map, and outputting a target recognition result; performing time phase division on the texture reconstruction image, judging a change type, and outputting a change detection result; and outputting a processing report including the enhanced image, the target list and change analysis. According to the method, noise suppression, correction enhancement, texture reconstruction, target recognition and change detection are integrated, the defect of fragmentation processing in the traditional technology is overcome, and the overall processing performance and the actual application adaptability are improved.

The invention provides a construction and use method of a potential diffusion model for SAR image super-resolution. The construction and use method comprises the steps of obtaining an original SAR image and inputting the original SAR image into a real degradation model to generate a degraded SAR image; inputting the degraded SAR image into an automatic encoder to generate a structure-enhanced submerged space feature map; and inputting the structure-enhanced latent space feature map and the degraded SAR image into a potential diffusion model to generate an SAR super-resolution image. The method has the beneficial effects that a two-stage training strategy is adopted, different optimization targets are focused in stages, the training efficiency is improved, and meanwhile, the learning ability of the model to SAR image features is enhanced; sAR imaging key degradation factors are comprehensively covered, so that a generated low-resolution sample is closer to a real scene, high-quality data support is provided for model training, and model learning is prevented from being separated from an actual degradation rule; sAR specific interference such as speckle noise is effectively simulated, and the anti-noise training effect of the model is enhanced.

The invention discloses a multi-modal ultrasonic image intelligent analysis system based on deep learning, particularly relates to the field of ultrasonic image intelligent analysis, and is used for solving the problems of speckle noise and displacement artifacts caused by sound beam attenuation during deep abdomen scanning of a bedside ultrasonic person with a relatively high in-vivo quality index. Partition noise reduction is guided through pixel noise distribution associated with depth and attenuation, and deep micro-textures are reserved; the color flow velocity and the elastic displacement are anchored through the noise reduction gray scale edge, three-mode sub-pixel alignment is completed, and cross-channel deformation and drift are eliminated; the self-attention fusion network synchronously extracts texture, speed and hardness features on the unified alignment surface to realize coupling of a complete biological structure and dynamic information; the texture residual error rate and the flow rate matching degree form complementary evaluation, noise distribution and alignment grid real-time self-adjustment are driven, frame-by-frame iteration convergence is achieved, and image definition and feature consistency are continuously improved.

The invention discloses an ultrasonic image denoising method based on variational mode decomposition and local space sparse fusion, which comprises the following steps of: firstly, adaptively optimizing key parameters of variational mode decomposition by using a grey wolf optimization algorithm to realize stable and efficient decomposition of an ultrasonic image; then, classifying the modal components according to the structural features of the modal components, and implementing differentiated denoising strategies for different types of modals to separate noise and reserve useful information; after modal reconstruction, a sparse expression method based on local space information is further introduced, according to the method, accurate boundary detection is carried out through gradient vector flow and gray scale proportion analysis, self-adaptive partitioning is carried out on an image according to boundary information, and finally sparse reconstruction is carried out through double dictionaries trained for different areas. According to the method, speckle noise in the ultrasonic image can be effectively suppressed, and meanwhile, the capability of keeping the edge and detail information of a tissue structure is remarkably improved, so that the ultrasonic image with higher quality is obtained.

The invention discloses a tracking optimization method for detecting a low-speed small unmanned aerial vehicle target by a single-photon laser radar, belongs to the technical field of optical detection and target tracking, and aims to solve the problem of unstable target imaging and tracking caused by speckle noise interference in the prior art. The invention provides a speckle noise suppression method based on a vibration emission optical fiber and a space-time dynamic kernel density estimation algorithm, and the method is combined with an improved mean shift-Kalman filtering algorithm to achieve the tracking optimization of a low-speed small unmanned aerial vehicle target, and obtains an echo signal through a single-photon laser radar. Constructing a three-dimensional data matrix and reconstructing distance and intensity images through space-time filtering and kernel density estimation; and then, in combination with gray level histogram probability estimation, similarity measurement and Mean Shift iteration, target area tracking is realized, and finally, a target position is predicted and output by using Kalman filtering. The method is suitable for the fields of long-distance unmanned aerial vehicle detection and monitoring, low-altitude security defense early warning, civil airspace management and control, military anti-unmanned aerial vehicle systems and the like.

The invention belongs to the technical field of image reconstruction and classification, and particularly relates to a radar image adaptive reconstruction and classification method based on intelligent hierarchical learning, and the method comprises the steps: obtaining an original radar image and radar platform parameters, and obtaining a standardized preprocessing image; outputting a high-fidelity reconstructed image through hierarchical processing; carrying out Sobel gradient calculation to obtain a gradient image and carrying out region division, carrying out differential dynamic window local enhancement, and then carrying out global gray balance and secondary filtering to obtain an enhanced and optimized image; carrying out local / global feature extraction, carrying out weighted fusion by utilizing semantic association, and outputting a multi-layer fusion feature map; outputting a high-confidence identification result by utilizing candidate region acquisition and difficulty evaluation, dynamic network switching identification and confidence closed-loop verification; and integrating the original radar image, the high-fidelity reconstructed image and the high-confidence identification result, and outputting a structured report. According to the method, the optimal balance of radar image speckle noise suppression and detail reservation can be realized.

The invention relates to the field of intelligent image enhancement display, in particular to an AI-based ultrasonic image pathological feature intelligent extraction and image enhancement display system, which comprises a preprocessing module, a feature extraction module, an image enhancement module, an interactive display module and a database, the preprocessing module acquires an original image, performs adaptive gray normalization and speckle noise suppression, and realizes accurate processing through an adaptive gray normalization network and a feature sensing speckle denoising network; the feature extraction module fuses multi-modal features by using a cross-modal attention mechanism, calculates attention matrix operation, outputs a pathological feature probability graph and generates a pathological feature mask; the image enhancement module drives a generator to enhance the contrast of a pathological area by taking the pathological feature mask as a condition through a conditional generative adversarial network; the interactive display module compares and plays images on a display screen; according to the invention, accurate extraction and image enhancement of ultrasonic image pathological features are realized.

The invention relates to an expressway ponding recognition method based on remote sensing data and deep learning. The expressway ponding recognition method comprises the following steps: collecting single-view complex image data in a synthetic aperture radar; carrying out denoising on the single-view complex image data, removing random speckle noise, and then carrying out smooth processing, geometric correction and radiometric calibration; geocoding is carried out in combination with DEM data, and imaging geometric distortion and position offset are removed; combining the remote sensing image, labeling a data label, and constructing a data set; constructing an image segmentation model based on a U-Net network; inputting an enhanced data set into the image segmentation model to carry out model training; and inputting a radar image needing ponding identification into the image segmentation model to carry out ponding identification in the expressway road domain range. The water body classification method based on the U-Net network is constructed and verified. By making a high-quality urban water body annotation data set and introducing various data enhancement strategies, the segmentation performance of the model in a complex remote sensing image is remarkably improved.

The invention discloses a retina OCT (Optical Coherence Tomography) image denoising method based on zero sample learning. The method comprises the following steps: generating a noise independent image pair for a single noisy image by adopting a CDIS (Coherent Discrete Identifier) The method comprises the following steps of: carrying out de-noising by using an MLFSnet network (AF-MSDSConv, Aap-LSM, SA, a reconstruction module); using RMSE symmetry and consistency joint loss zero sample training; in the reasoning stage, an original image is directly input, and a same-resolution de-noising result is output. The speckle noise can be significantly suppressed and the layered structure of the retina can be retained without noise-free true values or pairwise data, so that the retina can be used while being shot.

Systems and methods are provided for performing Doppler (phase-sensitive) optical coherence tomographic vibrometry measurements involving the vibrometric response of a rigid structure to an acoustic stimulus, such that the impact of speckle noise on a calculated vibrometric measure is reduced. Detected signals are processed to determine a depth range associated with the rigid structure, and to provide sampled time-dependent phase data encoded with the vibratory response of the rigid structure during application of the acoustic stimulus. The sampled time-dependent phase profile is processed to generate a vibrometric measure characterizing the rigid structure, based on non-uniformly-weighted contributions from one or more subregions within the depth region associated with the rigid structure (and optionally one or more time windows). The contribution from each subregion may be dependent on a depth-domain intensity measure associated with the subregion, where each subregion may be sufficiently small to encompass speckle-induced spatially-dependent changes.

The invention relates to the technical field of remote sensing monitoring, in particular to a method for monitoring the coastal reclamation condition through satellite remote sensing. The method comprises the following steps: acquiring SAR image data of a pair of target monitoring coast areas, and carrying out speckle noise suppression and distance Doppler terrain correction processing; according to the method, pixels in a main image and an auxiliary image in a deformation area in processed SAR image pair data are obtained, the coherence coefficient of the deformation area is calculated, and the deformation area within a set time range value is judged to be a candidate sea reclamation change suspected area according to the coherence coefficient of the deformation area and a coherence coefficient threshold value of the deformation area. According to the method, the candidate suspected area of the sea reclamation change is determined through the coherence coefficient focused on the deformation area, blind search of the whole monitoring range is avoided, the specific position of the sea reclamation change is accurately positioned, so that the possibility of misjudgment is reduced, meanwhile, the coherence coefficient focused on the deformation area can quickly exclude most deformation areas without change, and the accuracy of the sea reclamation change is improved. And the monitoring efficiency is improved.

The invention discloses a near-field electromagnetic wave imaging multi-mode noise cooperative suppression method, relates to the technical field of electromagnetic wave imaging, and aims to solve the problems that the suppression effect on speckle, stripe and Gaussian mixture noise is poor and a target structure is easy to lose in the prior art. According to the method, a complex field speckle suppression-multidirectional fringe separation-cross-channel Gaussian suppression three-stage cooperation scheme is adopted, and firstly, the amplitude and phase of a complex field image are processed through an adaptive threshold value to remove speckle noise; respectively executing horizontal / vertical ADOM filtering on the real part and the imaginary part of the complex field to eliminate stripe noise; and finally, Gaussian noise is suppressed by combining BM3D filtering and three-dimensional transform domain optimization, and robustness is improved through multi-frequency point fusion. Experiments show that the SSIM of the method is improved by 21.7% compared with that of a traditional method, the GSSIM and the PSNR are optimal, the structural features of the target can be reserved in a strong noise environment, and the method is suitable for scenes such as defect detection of near-field electromagnetic wave synthetic aperture imaging.

The invention discloses a physical constraint deep learning forest biomass estimation method and system. The method specifically comprises the steps of satellite-borne laser radar GEDI footprint preprocessing and quality screening; sAR radiometric calibration, terrain correction and speckle noise filtering are carried out; optical data radiation correction, atmospheric correction and geometric correction; multi-source remote sensing data space-time alignment and feature variable extraction; constructing a deep learning model fused with physical constraints; feature optimization and model reconstruction are carried out based on space-time interpretability analysis (SHAP); performing comparison and precision verification on the reconstructed final model; and generating a high-resolution annual forest biomass map based on the optimal deep learning model. The estimation result obtained by the method provided by the invention has relatively high precision and accuracy, and scientific data support and decision suggestions are provided for regional forest resource management and carbon sink capability evaluation.

The invention discloses a diffusion-guided image cross-modal matching method. The method comprises the following steps: firstly, preprocessing an optical image and an SAR image to be matched; obtaining a coarse-to-fine multi-scale feature pair of the optical image and the SAR image and a multi-scale enhancement feature of the SAR image by using a diffusion gating feature pyramid network; on the thickest scale, an optical flow field probability distribution model is constructed through a Gaussian process matching module, and an initial optical flow field and a confidence map are output in combination with a Transform decoder; then, a multi-scale convolution refining network is adopted, displacement embedding and local correlation calculation are combined, step-by-step optimization of the optical flow field and confidence is achieved through iterative updating, and finally a fine matching result is output. According to the method, speckle noise in the SAR image is effectively suppressed and geometric distortion is compensated through feature enhancement guided by the diffusion model, Gaussian process modeling, Transform decoder prediction and a multi-scale progressive refinement strategy, and the robustness and precision of cross-modal image matching in a complex scene are remarkably improved.

This application relates to a method, apparatus, device, and medium for few-sample SAR target recognition. The method includes: dividing SAR samples into support and query sets based on a meta-learning framework; converting the original SAR image into a semantic map and an attribute scattering center topology map; extracting three types of features and calculating class prototypes through three parallel feature branches; obtaining the confidence score of each branch using the Softmax function; assigning teacher and student branches according to the confidence score; minimizing KL divergence to achieve dynamic knowledge transfer; weighted fusion of predicted log odds followed by Softmax output; and backpropagation to optimize the network. This method, through multi-prototype fusion and interactive distillation, mitigates the problems of speckle noise and high intra-class variance in SAR images, maintaining high robustness and generalization ability even in extreme data-scarce scenarios.

The invention relates to a near-to-eye display optical system based on a laser light source and an LCOS modulation module, and belongs to the technical field of near-to-eye display. The invention aims to solve the technical problems that the existing near-to-eye display scheme is difficult to overcome speckle noise, chromatic dispersion, low luminous efficiency, overlarge size and the like at the same time. The system sequentially comprises a multi-mode laser light source module used for emitting low-initial speckle contrast imaging laser beams; the optical preprocessing module comprises a collimating lens group, a beam expanding prism and a phase perturber; the LCOS modulation module is used for simultaneously carrying out amplitude modulation on the light beam according to the image signal so as to load an image and carry out phase modulation so as to adjust wavefront, and reflecting and outputting an image light beam; the coupling module is used for carrying out light path coupling on the image light beam; and the near-to-eye display module is used for transmitting and projecting an image to human eyes. The system is mainly suitable for augmented reality glasses and other head-mounted display devices, and can realize high-image-quality and miniaturized near-to-eye imaging.

The invention discloses an SAR (Synthetic Aperture Radar) image change detection method combining convolution and mixed attention. The SAR image change detection method comprises the following implementation steps of: firstly, generating a difference chart for two SAR images by using a composite neighborhood intensity difference method; then, a hierarchical FCM clustering algorithm is used for carrying out pre-classification processing on the difference image, a pseudo label matrix is generated, variable and invariable high-probability sample pixels in pseudo label pixels are selected, spatial positions of the pixels are extracted, and on the pixels of the corresponding spatial positions of the two original SAR images, a pseudo label matrix is generated; pixel blocks with the pixel points as the centers are taken as a training set, and pixel blocks with all the pixel points as the centers are extracted from the two original SAR images to serve as a test set; and then training a neural network combining convolution and mixed attention by using the training sample set, and then carrying out change detection analysis on a test set by using the trained network to generate a final change detection result graph. The method has clear advantages in the aspects of SAR speckle noise suppression and change detection precision.

This invention provides a method for selecting strong scattering points in space-based ISAL systems by incorporating multi-dimensional features, relating to the field of ISAL image processing. This invention employs a multi-stage fusion processing strategy to form a complete link from image preprocessing to the final output of strong scattering points: pixel replication and extension are used to process convolutional boundary regions, significantly reducing boundary amplitude distortion caused by traditional zero-padding or periodic extension, and improving the signal-to-noise ratio of edge regions in the filtered amplitude map; the dynamic threshold generation process works synergistically with the preceding Gaussian kernel convolution preprocessing to suppress speckle noise while preserving the true amplitude characteristics of weakly scattering targets; amplitude-weighted averaging is used to calculate the centroid coordinates within the connected domain, overcoming the energy diffusion problem caused by the ISAL system's point spread function; transient interference points are eliminated based on the peak-to-mean ratio of multi-frame amplitude sequences, and the spatial distribution characteristics of the centroid position's spatial amplitude value and density constraints are combined to ensure that the output set of preferred scattering points simultaneously satisfies temporal stability, spatial saliency, and uniform distribution.

This invention belongs to the field of automatic target recognition security technology for synthetic aperture radar (SAR), and discloses a method and application for generating adversarial examples for SAR images. The method includes: generating a more stable gradient prediction look-ahead direction by combining the gradient and cumulative momentum of the current data point to reduce the influence of speckle noise; then, effectively sampling the area around the data point and adjusting its weights according to its confidence score to reduce the negative impact of semantic inconsistency of the sampled data points and further stabilize the gradient update. This invention considers the problem that SAR images are prone to excessive changes in gradient direction during gradient calculation due to coherent noise. By combining the gradient of the current data point and cumulative momentum to stabilize the gradient prediction direction, and by sampling the surrounding data points through confidence scores to achieve a more stable adversarial example update direction, it can well fit the characteristics of SAR data, thereby generating high-quality adversarial examples in real-world black-box scenarios.

The application discloses a near-field electromagnetic wave imaging multi-modal noise cooperative suppression method, relates to the technical field of electromagnetic wave imaging, and aims to solve the problems of poor suppression effect of spot noise, stripe noise and Gaussian mixed noise and easy loss of target structure in the prior art. The method adopts a three-stage cooperative scheme of "complex domain spot suppression-multi-direction stripe separation-cross-channel Gaussian suppression", first removes the spot noise by adaptively processing the amplitude and phase of the complex domain image through a threshold value; then eliminates the stripe noise by performing horizontal / vertical ADOM filtering on the real part and the imaginary part of the complex domain, respectively; finally, combines BM3D filtering and three-dimensional transform domain optimization to suppress Gaussian noise, and improves the robustness through multi-frequency point fusion. Experiments show that the SSIM of the method is improved by 21.7% compared with the traditional method, the GSSIM and PSNR are optimal, the target structure features can be reserved in a strong noise environment, and the method is suitable for defect detection and other scenes of near-field electromagnetic wave synthetic aperture imaging.

The invention provides a CNN and SAM fusion-based ultrasonic medical image segmentation method CASAM, and aims to construct a CNN and SAM ViT double-branch collaborative framework for the problems of low contrast, strong speckle noise, fuzzy anatomical boundary, complex lesion form, difficulty in small target segmentation and the like of an ultrasonic image. A dual-channel feature fusion module is arranged in the CNN branch to extract fine-grained local texture and edge information; local features are injected into ViT branches through a cross-branch low-rank adapter, and the characterization capability of small-scale details and key boundaries is enhanced under low parameter overhead; a medical cue adapter is introduced into a ViT branch, semantic and local clues are fused through image embedding, lesion boundary expression is enhanced in combination with edge information, and therefore lesion positioning and segmentation precision is improved. Experimental results show that higher precision and stronger generalization are achieved in multiple types of ultrasonic segmentation tasks, GPU video memory and computing overhead are reduced, and the method is suitable for resource-limited clinical terminal deployment.

The invention discloses a laser decoherence and holographic display speckle noise suppression system and method based on controllable fog scattering. The system comprises a laser light source module, an optical shaping module, a fog scattering module and a holographic display module which are arranged in sequence. The fog scattering module generates a scattering medium layer composed of micron-sized fog drops in an ultrasonic atomization or airflow disturbance mode, so that the space and time coherence of the collimated laser is reduced under the action of multiple times of scattering, and therefore speckle noise in holographic display imaging is effectively restrained. By adjusting the particle size distribution, the concentration and the disturbance frequency of the fog drops, continuous controllability of the decoherence intensity can be achieved, and the imaging brightness and the imaging definition are both considered. The system is simple in structure and high in compatibility, and can be directly embedded into an existing computer-generated holographic display system without an additional optical modulation device; and meanwhile, the method has the advantages of low cost, high real-time performance, suitability for a multiband laser system and the like, and is suitable for the optical fields of holographic display, laser projection, interference measurement and the like.

The application provides a SAR image target detection method based on SAR image sparse enhancement and Bayesian saliency detection; comprising: step one: inputting a matched filter SAR image, obtaining initial values of regularization parameters of a SAR sparse enhancement method by calculating gray statistical characteristics of the image; step two: solving a convolution sparse feature enhancement algorithm model based on L1 norm by using an alternating direction multiplier method, and obtaining a SAR sparse enhancement image; step three: calculating a Bayesian saliency map of the SAR sparse enhancement image, binarizing the saliency map by using a threshold detection method, and realizing final target detection. The method can reduce SAR speckle noise by using a sparse image enhancement method, reduce background gray values, highlight target regions by using a SAR image saliency detection method, further reduce background pixel values, and finally realize efficient detection of ground and sea surface targets of interest.

The application discloses a kind of self-adapting shadow enhancement method and device of smoke-passed infrared digital holographic reconstructed image, belong to image processing technical field.The application proposes a new image processing algorithm to enhance the shadow area caused by smoke-passed infrared digital holographic reconstructed image smoke sheltering, the algorithm is first using bilateral filter to remove noise to image, in combination with binaryzation, edge extraction, morphological algorithm divides shadow area, and in combination with bright area to achieve the purpose of enhancing the brightness of shadow area.The method disclosed in the application solves the problems of severe speckle noise, obvious shadow area, uneven brightness, missing object detail information and other problems of smoke-passed infrared holographic reconstructed image, can effectively enhance the image quality of smoke-passed infrared holographic reconstructed image, has significant effect in the aspects of shadow area enhancement and speckle noise removal, and belongs to the field of image processing technology.

The invention provides a low-speckle noise holographic 3D display method based on an improved random pixel separation method. According to the method, a 3D object is regarded as being composed of 2D objects with different depths, firstly, intensity information of the 2D objects is extracted to obtain an original image, and random pixel indexes of complementary checkerboards are extracted to generate a group of random masks; and then, superposing the random mask and the initial image to obtain a group of sub-images, applying a uniform phase to form an object plane complex amplitude, and calculating a holographic plane complex amplitude based on an angular spectrum diffraction principle. And then converting the complex amplitude into a phase hologram by adopting a double-phase coding method. And finally, the phase holograms are sequentially loaded to the spatial light modulator in a time division multiplexing mode, and a holographic 3D reproductive image with low speckle noise is reconstructed.

The application provides a SAR image denoising method based on self-supervised learning, comprising the following steps: constructing a SAR noise image model; constructing a minimum empirical loss function L; designing an adjacent co-directional subsampler; taking a noise image pair as input and target, modifying an empirical risk minimization task, and constructing a self-supervised SAR image denoising network; designing a multi-feature loss function; using the above method, training a SAR-CNN denoising network; transforming a SAR image to be denoised into a log domain and inputting the SAR image into the network for denoising, and then obtaining an output image through inverse log transformation. Compared with the prior art, the application can be migrated to any existing denoising network, realize self-supervised denoising in the absence of a clean target, significantly suppress speckle noise, reliably retain image texture details and other features, and solve the problem of poor practicability of current speckle data and real SAR image self-supervised denoising.

The invention provides a change detection method based on deformable convolution and state space feature reconstruction. The method is used for an SAR image change detection task, irregular change characteristics caused by geometric deformation and noise interference in an SAR image are adaptively captured through a deformable convolution encoder, a long-range global dependency relationship is established by using an AttnGate Mama decoder based on a state space model, and interference of speckle noise and registration errors is effectively overcome. Meanwhile, efficient fusion of local details and global semantics is realized through a designed multi-scale adaptive fusion gating mechanism and a forced binarization loss function, and the positioning precision of the boundary of the change region is remarkably improved. Efficient, robust and accurate SAR image change detection is realized.