Change segmentation method, system and medium based on learnable dct and frequency band division

Abstract

The application discloses a change segmentation method and system based on a learnable DCT and frequency band division and a medium, relates to the technical field of image processing, takes an improved AFB_DCT_FEM model as a core, connects a learnable two-dimensional DCT transformation network, an adaptive frequency band division AFB network and a Band Router dynamic weighting network into an overall framework that can be end-to-end back propagated, while maintaining the DCT energy concentration characteristics, endows the network with three abilities of frequency basis on-demand tuning, dynamic frequency band selection and automatic weight estimation, replaces the substructure of an existing image change region segmentation network in an end-to-end manner, realizes efficient extraction of double-time-phase image difference features, and significantly improves the fineness and robustness of image change region segmentation.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a variation segmentation method, system, and medium based on learnable DCT and frequency band division. Background Technology

[0002] Image change region segmentation is a key technology that uses multi-temporal high-resolution images (generally "temporal phase 1" and "temporal phase 2" of the same area) to determine the presence and type of change of surface targets in the temporal dimension. Its results directly serve numerous fields such as land resource monitoring, urban expansion assessment, disaster damage statistics, agricultural yield estimation, and military reconnaissance. Due to the characteristics of remote sensing images, such as high resolution, complex scenes, and drastic fluctuations in imaging conditions with seasons and climates, accurately and robustly extracting change regions under conditions of noise interference, lighting differences, changes in observation angle, and even inconsistent resolution is one of the technical challenges that has garnered significant attention from scientific research and industry.

[0003] One of the more representative methods in recent years is to cascade a frequency domain enhancement module (FEM) and a spatial reconstruction module (SRM). The FEM performs a two-dimensional discrete cosine transform (2-DDCT) on each image and selects "a fixed number of high-frequency coefficient channels" for subtraction and then superimposes them to emphasize fine-grained changes. The SRM performs multi-scale cross-fusion in the spatial domain and uses residual upsampling to recover texture. Finally, dual-domain feature stitching and decoding are performed to output a change map. This scheme has been validated for F1 enhancement on public datasets such as LEVIR-CD, WHU-CD, and S2Looking. The visualization results show that it is more sensitive to fine lines on roads and the demolition and reconstruction of building roofs. However, its technical shortcomings are also quite prominent: for example, the selection of high-frequency channels is fixed, and the human-imposed threshold leads to large fluctuations in the transfer effect between different datasets; the Discrete Cosine Transform (DCT) basis functions are still static, without learnable parameters, and cannot reshape the frequency domain basis according to the downstream loss; there is a lack of frequency band granular gating, all high frequencies are treated the same, and noise frequencies are also amplified; the memory usage is relatively high, and it is necessary to maintain the DCT coefficient tensor and spatial features at the same time, which limits the inference resolution; most models choose simple concatenation or element-wise weighting, and use cross attention, but lack true cross-domain interaction, resulting in high frequency and spatial texture being independent, making it difficult to make full use of complementary information. Summary of the Invention

[0004] To address the challenges of traditional image variation region segmentation methods, such as the inability of DCT basis functions to fully extract frequency domain information, the inability of DCT basis functions to reshape the frequency domain basis based on downstream losses, and the difficulty in fully utilizing complementary information due to the independence of high frequencies and spatial textures, this application provides a variation segmentation method, system, and medium based on learnable DCT and frequency band division. Using an improved AFB_DCT_FEM model as its core, a learnable 2D DCT transform network, an adaptive frequency band division AFB network, and a Band Router dynamic weighting network are connected in a unified framework capable of end-to-end backpropagation. While maintaining the energy concentration characteristics of DCT, this approach endows the network with the triple capabilities of on-demand frequency basis tuning, dynamic frequency band selection, and automatic weight estimation, thereby significantly improving the precision and robustness of image variation region segmentation.

[0005] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0006] This solution provides a variation segmentation method based on learnable DCT and frequency band division, including:

[0007] Acquire a first-phase image and a second-phase image of the same target area, and preprocess the first-phase image and the second-phase image;

[0008] The preprocessed first and second temporal images are input into the AFB_DCT_FEM model to obtain the block-level DCT coefficient tensor, and differential features are generated through adaptive frequency band division and routing weighting.

[0009] The preprocessed first and second temporal images are fed into the spatial encoder in parallel to extract multi-scale spatial features.

[0010] The difference features are fused with the multi-scale spatial features to obtain a pixel-level variation mask.

[0011] A further optimized solution is that the preprocessing includes:

[0012] The first and second temporal images are corrected, including radiometric correction and geometric correction.

[0013] The corrected first and second phase images are cropped to a target block size, which is a size divisible by a block of size B×B.

[0014] A further optimized solution is that the method for obtaining the differential features includes:

[0015] Based on a learnable 2D DCT transform network, a cross-domain fine-tuning strategy is used to perform block-level frequency domain mapping on either the first or second temporal image, resulting in a mapped spectrum containing multiple frequency points; the DCT convolution kernel of the learnable 2D DCT transform network is a trainable DCT convolution kernel, and the DCT basis functions are:

[0016] ; ;

[0017] Where x and y represent the spatial coordinate indices within the block; B s The convolution block size is represented by u and v; the frequency domain coordinates are represented by u and v. The length weight of the k-th spectrum segment is indicated; k = u or v; cos[] represents the cosine function;

[0018] The adaptive frequency band division AFB network segments the mapped spectrum according to the trainable frequency band threshold.

[0019] The Band Router dynamic weighted network scores and weights each spectrum segment obtained from the segmentation, and then upsamples to obtain differential features.

[0020] A further optimization scheme is proposed, in which the cross-domain fine-tuning strategy includes:

[0021] Initialize the DCT convolution kernel, construct the adaptive attenuation coefficient matrix of the target image sensor, and scale the DCT convolution kernel based on the adaptive attenuation coefficient matrix after loading the pre-trained weights.

[0022] The spatial resolution of the target image and the baseline spatial resolution during the pre-training stage of the learnable 2D DCT transform network are obtained. The size of the convolutional block is determined by combining the baseline spatial resolution and the spatial resolution of the target image to perform multiple interpolation scaling operations on the DCT convolutional kernel. The size of the convolutional block satisfies the following:

[0023] ;

[0024] in, Indicates the size of the scaled convolutional block; r base Indicates the reference spatial resolution; r target B represents the spatial resolution of the target image; s Indicates the size of the convolution block;

[0025] The gain function for each frequency point is determined based on the dark channel prior method, and the DCT convolution kernel is scaled according to the frequency point based on the gain function.

[0026] A further optimized scheme involves determining the gain function for each frequency point using the dark channel prior method, and scaling the DCT convolution kernel by frequency point based on the gain function, including the following method:

[0027] Atmospheric light transmittance t was estimated based on the dark channel prior method;

[0028] The gain function for each frequency point is calculated based on the following formula:

[0029] ;

[0030] ;

[0031] in, B represents the frequency radius, k represents the hyperparameter controlling the gain intensity; s Indicates the size of the convolution block; exp[*] represents the exponentiation operation; Represents the gain function; u and v represent frequency domain coordinates;

[0032] Scaling the DCT convolution kernel by frequency point based on the gain function at each frequency point:

[0033] ;

[0034] in, This represents the final learnable DCT convolutional weights that take effect. This represents the learnable DCT convolution weights adapted after training.

[0035] A further optimized scheme is that the adaptive frequency band partitioning AFB network segments the mapped spectrum according to a trainable frequency band threshold; including the following method:

[0036] Configure the number of spectrum segments K, and the set of edge thresholds for the spectrum segments is as follows: in, The threshold value for the (K-1)th spectral segment satisfies ;

[0037] The unsorted learnable scalars are identified, and the length weights of each spectral segment are obtained by combining the softmax function. The ordered boundary thresholds are then obtained based on prefix accumulation.

[0038] For each spectral band, a soft mask is generated based on the double sigmoid difference function.

[0039] A further optimization scheme involves using a Band Router-based dynamic weighted network to score and weight each segmented spectrum band, then fusing them together, and finally upsampling to obtain differential features; including the following methods:

[0040] Obtain the frequency band characteristics of K frequency bands ;

[0041] For the k-th frequency band feature F kThe frequency band feature vector P is extracted through lightweight convolution. k The frequency band feature vector P k It characterizes local frequency band information and global statistical information;

[0042] All frequency band feature vectors are concatenated, and the weights of each frequency band are predicted through a fully connected layer. The weighted fusion is then used to output differential features.

[0043] A further optimization scheme is proposed, in which the loss function of the Band Router dynamically weighted network is:

[0044] ;

[0045] ;

[0046] in, This represents the loss function of the Band Router dynamically weighted network. This represents the average activation weight of the k-th spectral segment within a batch, where N represents the batch size. This represents the prediction weight of the nth sample in the kth spectral segment.

[0047] This solution also provides a variation segmentation system based on learnable DCT and frequency band division, for implementing the aforementioned variation segmentation method based on learnable DCT and frequency band division. The system includes:

[0048] The acquisition module is used to acquire first and second phase images of the same target area;

[0049] The preprocessing module is used to preprocess the first temporal image and the second temporal image;

[0050] The first feature extraction module is used to input the preprocessed first and second temporal images into the AFB_DCT_FEM module to obtain block-level DCT coefficient tensors, and generate differential features through adaptive frequency band division and routing weighting.

[0051] The second feature extraction module is used to send the preprocessed first temporal image and the second temporal image into the spatial encoder in parallel to extract multi-scale spatial features.

[0052] The fusion output module is used to fuse the differential features with the multi-scale spatial features to obtain a pixel-level variation mask.

[0053] This solution also provides a computer-readable medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, can implement the variation segmentation method based on learnable DCT and frequency band division as described above.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. This application provides a variation segmentation method, system, and medium based on learnable DCT and frequency band division; with an improved AFB_DCT_FEM model as the core, a learnable two-dimensional DCT transform network, an adaptive frequency band division AFB network, and a BandRouter dynamic weighting network are connected in series to form an overall framework with end-to-end backpropagation. While maintaining the energy concentration characteristics of DCT, the network is endowed with the triple capabilities of on-demand frequency base optimization, dynamic frequency band selection, and automatic weight estimation; and the substructure of the existing image variation region segmentation network is replaced in an end-to-end manner to achieve efficient extraction of differential features of dual-temporal images, significantly improving the accuracy and robustness of image variation region segmentation.

[0056] 2. This application provides a variation segmentation method, system, and medium based on learnable DCT and frequency band division; its learnable two-dimensional DCT transform network explicitly solidifies the two-dimensional DCT basis functions into convolution kernels and sets the convolution weights as trainable parameters, realizing learnable frequency domain transform parameters; combined with orthogonal constraint loss, it preserves the energy concentration and interpretability of the basis, taking into account the high-frequency focusing characteristics of DCT and the differentiability advantage of convolutional networks; it retains the good energy concentration characteristics of DCT, and allows the network to automatically find the optimal frequency domain representation that better matches the actual ground features and noise distribution during training, thereby significantly improving the separability of subtle variations.

[0057] 3. This application provides a variation segmentation method, system, and medium based on learnable DCT and frequency band division; during the adaptive frequency band division process, a connected and differentiable soft frequency band mask is generated based on the "double Sigmoid" function to avoid gradient discontinuity caused by hard thresholding; at the same time, the adaptive frequency band division AFB network dynamically adjusts the segmentation threshold and frequency band width according to the statistical energy proportion and gradient sensitivity of each frequency point during training, realizing data-driven frequency domain subspace recombination, which not only automatically suppresses noise frequency bands and retains the frequency bands most sensitive to changes, but also reduces redundant computation of subsequent networks.

[0058] 4. This application provides a variation segmentation method, system and medium based on learnable DCT and frequency band division; its BandRouter dynamic weighted network assigns weights to each adaptive frequency band and selects the most suitable feature extraction path to form a frequency band-specific deep sub-network, which strengthens the information complementarity between high frequency and low frequency, while reducing frequency band redundancy interference, making the model more robust under multi-scale and cross-seasonal imaging conditions. Attached Figure Description

[0059] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a schematic diagram of the process for a segmentation method based on learnable DCT and frequency band division.

[0061] Figure 2 This is a schematic diagram illustrating the principle of a segmentation method based on learnable DCT and frequency band division.

[0062] Figure 3 This is a schematic diagram of the AFB_DCT_FEM model structure;

[0063] Figure 4 This is a schematic diagram of a segmentation system based on learnable DCT and frequency band division.

[0064] Figure 5 This is a diagram showing the comparison results of early-stage images, later-stage images, change ground truth, and model output. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0066] Traditional image segmentation methods struggle to fully extract frequency domain information, the Discrete Cosine Transform (DCT) basis function cannot reconstruct the frequency domain basis based on downstream loss, and the independence of high frequencies and spatial texture makes it difficult to fully utilize complementary information. Therefore, this solution provides the following embodiments to address these technical problems:

[0067] This embodiment provides a variation segmentation method based on learnable DCT and frequency band division, such as Figure 1 and Figure 2 As shown, the method includes:

[0068] Step 1: Acquire a first temporal image I1 and a second temporal image I2 of the same target area, and preprocess the first temporal image I1 and the second temporal image I2.

[0069] The preprocessing described in step one includes:

[0070] The first and second temporal images are corrected, including radiometric correction and geometric correction.

[0071] The corrected first and second phase images are cropped to a target block size, which is a size divisible by a block of size B×B.

[0072] Step two involves inputting the preprocessed first and second temporal images into the AFB_DCT_FEM model to obtain block-level DCT coefficient tensors, and generating differential features through adaptive frequency band division and routing weighting; such as Figure 3 As shown, the specific methods for obtaining differential features include:

[0073] S21, based on a learnable two-dimensional DCT transform network, a cross-domain fine-tuning strategy is used to perform block-level frequency domain mapping on either the first or second temporal image, resulting in a mapped spectrum containing multiple frequency points; the DCT convolution kernel of the learnable two-dimensional DCT transform network is a trainable DCT convolution kernel, and the DCT basis function is:

[0074] ; ;

[0075] Where x and y represent the spatial coordinate indices within the block; B s The convolution block size is represented by u and v; the frequency domain coordinates are represented by u and v. The length weight of the k-th spectrum segment is indicated; k = u or v; cos[] represents the cosine function;

[0076] Parameterize the DCT basis functions into trainable convolutional kernels W. u，v And by adding orthogonal regularization, we have:

[0077] ;

[0078] in, The loss function of a learnable 2D DCT transform network is represented by ; W represents the basis matrix formed by stacked convolutional kernels; I represents the identity matrix. The Frobenius norm is represented by T; T represents the transpose.

[0079] To ensure that the basis functions remain approximately orthogonal during learning and to prevent high- and low-frequency leakage, in a specific implementation, trainable DCT convolution kernels can be generated by the dct_kernels function using the nn.Parameter class.

[0080] In step S21, the cross-domain fine-tuning strategy includes:

[0081] Different remote sensors (such as optical satellites, multispectral payloads, and UAVs) exhibit differences in their imaging modulation transfer function (MTF); therefore, this scheme employs a cross-domain fine-tuning strategy, including sensor domain adaptation and related methods.

[0082] Initialize the DCT convolution kernel as Where c represents the channel index; u and v represent the horizontal and vertical frequency indices; for a target image sensor, its modulation transfer function can be approximated by the following form:

[0083] ;

[0084] in, The normalized frequency is represented by a0, a1, and a2, which are the coefficients obtained by fitting experimental calibration data.

[0085] Construct the adaptive attenuation coefficient matrix of the target image sensor After loading the pre-trained weights, the DCT convolution kernel is scaled based on the adaptive decay coefficient matrix:

[0086] ;

[0087] in, This represents the final learnable DCT convolutional weights that take effect. This represents the learnable DCT convolution weights that are adapted after training.

[0088] This operation is performed once after the pre-trained weights are loaded and does not participate in gradient updates;

[0089] When the input image resolution differs from that of the model during pre-training, to maintain consistency in frequency resolution, the receptive field of the DCT convolutional kernel needs to be scaled. This includes methods such as obtaining the spatial resolution of the target image and the baseline spatial resolution of the learnable 2D DCT transform network during pre-training; determining the size of the convolutional block by combining the baseline spatial resolution and the spatial resolution of the target image; performing multiple interpolation scaling on the DCT convolutional kernel; specifically, performing three interpolation scaling operations; and performing energy normalization to ensure consistent frequency domain response of the convolution operation. The adjusted DCT convolutional kernel will be used as initial weights for further optimization during fine-tuning training in the target domain. The size of the convolutional block satisfies the following:

[0090] ;

[0091] in, Indicates the size of the convolution block; r base Indicates the reference spatial resolution; r target B represents the spatial resolution of the target image; s Indicates the size of the convolution block; This indicates rounding up to the nearest integer.

[0092] To eliminate the interference of ambient lighting changes (such as shadows, haze, and atmospheric scattering) on ​​the frequency distribution, this scheme proposes a method for scaling the DCT convolution kernel by frequency point, specifically including the following steps:

[0093] The gain function for each frequency point is determined based on the dark channel prior method, and the DCT convolution kernel is scaled according to the frequency point based on the gain function; specifically:

[0094] Atmospheric light transmittance t was estimated based on the dark channel prior method;

[0095] The gain function for each frequency point is calculated based on the following formula:

[0096] ;

[0097] ;

[0098] in, B represents the frequency radius, k represents the hyperparameter controlling the gain intensity; s Indicates the size of the convolution block; exp[*] represents the exponentiation operation; The gain function is represented by u and v, which represent frequency domain coordinates.

[0099] Scaling the DCT convolution kernel by frequency point based on the gain function at each frequency point:

[0100] ;

[0101] in, This represents the final learnable DCT convolutional weights that take effect. This represents the learnable DCT convolution weights adapted after training.

[0102] This method can effectively recover mid-to-high frequency energy and improve the ability to detect change regions in low-contrast environments.

[0103] First, the AFB_DCT_FEM model receives feature maps of size B×C×H×W (B represents the number of samples in a batch; C represents the number of channels in the intermediate features; H represents the spatial height of the feature map; W represents the spatial width of the feature map), and performs block-level frequency domain mapping on them using a learnable 2D DCT transform network. Specifically, the DCT convolution kernels are preset as orthogonal basis functions and declared as trainable parameters, allowing the network to make fine adjustments to the basis to fit the data distribution while maintaining the "energy concentration" characteristic. The convolution stride (the size of the convolution block) is B. s (Default 8), depthwise separable convolution is implemented using the nn.Conv2d class function, with each input channel corresponding to... The output tensor has 64 output channels. The initial values ​​of the convolution kernels are generated by two-dimensional discrete cosine transform orthogonal basis in the (u, v) direction, and then participate in backpropagation for update. Therefore, the resolution of the output tensor is reduced to (H / 8, W / 8) in blocks, and the number of channels is expanded to 64C, which is equivalent to splitting each original channel into 64 frequency channels.

[0104] Traditional DCT layer coefficients are fixed and cannot tolerate spectral drift introduced by scene differences. This scheme achieves data-driven adjustment and improves generalization by combining orthogonal initialization with constraint regularization while preserving the energy concentration of DCT. Through parameter fine-tuning and learnability, the DCT convolution kernel can be finely tuned to adapt to different sensors, image resolutions, lighting and atmospheric conditions. Under the same training epochs, the accuracy is improved compared to fixed DCT, and the gradient is more concentrated.

[0105] In remote sensing change detection tasks, the spectral energy distribution of input images varies significantly depending on the land cover category. For example, urban building rooftops typically exhibit strong signals in the mid-to-high frequency range, while regular textures such as farmland are mainly distributed in the mid-to-low frequency range, and water bodies or bare land are dominated by low-frequency flat components. Traditional methods use fixed frequency band boundaries, which cannot simultaneously accommodate multiple land cover scenarios, easily leading to problems such as high-frequency noise amplification, loss of low-frequency features, or energy leakage. Therefore, this solution proposes an adaptive frequency band division method, allowing the network to automatically learn and dynamically adjust the optimal frequency band boundaries according to the specific task during training, thereby improving segmentation performance and generalization ability. Considering that frequency features are easily damaged in the early stages of training, this solution introduces a phased training scheduling strategy to prevent overfitting of low-frequency channels, as detailed in step S22.

[0106] S22, the AFB network based on adaptive frequency band segmentation performs frequency segmentation on the mapped spectrum according to the trainable frequency band threshold; step S22 specifically includes the following method:

[0107] Configure the number of spectrum segments K, and the set of edge thresholds for the spectrum segments is as follows: ;in, The threshold value for the (K-1)th spectral segment satisfies ;

[0108] To determine unsorted learnable scalars, this embodiment uses β. k For unsorted learnable scalars, k = 1, 2, ..., K-1; the length weights of each spectrum segment are obtained using the softmax function, and the ordered boundary thresholds are obtained based on prefix accumulation; specifically, the length weight of the k-th spectrum segment is:

[0109] ;

[0110] Where, β j Let j represent the j-th unsorted learnable scalar; Indicates the weight of the length of the k-th spectrum segment;

[0111] The ordered boundary threshold is obtained by prefix accumulation as follows:

[0112] ;

[0113] Therefore, it is automatically satisfied. This method ensures that each threshold increases monotonically, effectively avoiding the training instability caused by traditional threshold crossing, and maintaining the differentiability of parameters and good gradient propagation.

[0114] For each spectral band, a soft mask M is generated based on the bi-Sigmoid difference function. k (r):

[0115] ;

[0116] in, Let θ0 represent the Sigmoid function; τ represents the boundary smoothing hyperparameter (usually taken as 0.05 to 0.2), θ0 = 0, θ0 = 0. K =1.

[0117] After obtaining the complete spectrum, the next step is the adaptive frequency band allocation stage, where the normalized radius r is pre-calculated for each frequency point:

[0118] ;

[0119] in, ; Indicates the diagonal radius;

[0120] The system learns (n-1) thresholds θ and registers them as constant vectors to ensure zero overhead during inference. The adaptive frequency band division AFB network holds a trainable threshold vector θ, which is forced to monotonically increase by first using the softmax function and then accumulating, naturally avoiding threshold "conflicts". Subsequently, a differentiable soft mask m is generated for each spectrum band using the "double sigmoid" function. k (r), obtain n frequency band features (frequency band tensors) F k The DCT coefficients are divided into K spectrum segments of low, medium and high. Each spectrum segment is a list of sub-tensors of the same size as the original DCT tensor. This soft mask keeps the gradient continuous near the threshold and allows the threshold to automatically migrate to the most discriminative radius range during training.

[0121] To accelerate network convergence and guide effective segmentation, the initial boundary thresholds can be uniformly set to 0.25, 0.5, and 0.75. During training, a frequency band sparsity regularization term is added to the main loss function, with the following constraints:

[0122] ;

[0123] in, This indicates the band sparsity regularization loss; Represents the regularity coefficient. This represents the L1 norm of the mask for the k-th spectral segment; this regularization helps to eliminate redundant frequency bands and improve the model's resolution and compression capabilities.

[0124] The masking function described above achieves a smooth transition at each frequency band boundary, ensuring that the segmentation operation is differentiable with respect to the network parameters, and is beneficial to the stability and generalization of model training.

[0125] Existing technologies often use methods such as "directly extracting the first K high-frequency coefficients" or "manually defining fixed frequency rings," lacking a differentiable and adaptive frequency band selection mechanism. This scheme uses a sigmoid dual-gating method to form a continuously differentiable soft mask, resulting in more stable gradients. To further improve feature separation capabilities in tasks such as segmenting changing regions in multi-scene images, this scheme employs dynamic weight fusion for feature components of different frequency bands after adaptive frequency band division. Since the contribution of each frequency band to the final decision is significantly affected by the input sample type (e.g., city, farmland, water body), static weighting or average fusion often leads to feature redundancy and decreased discriminative power. Therefore, this application proposes a Band Router dynamic weighting network to achieve sample-level dynamic weighting of frequency-divided features, thereby improving the system's scene adaptability and inference efficiency. Specifically, this is step S23.

[0126] S23, based on the Band Router dynamic weighted network, scores and weights the obtained spectrum segments, and finally upsamples to obtain differential features; step S23 specifically includes the following methods:

[0127] S231, Obtain the frequency band characteristics of K frequency bands. ;

[0128] S232, for the k-th frequency band feature F k The frequency band feature vector P is extracted through lightweight convolution. k The frequency band feature vector P k It characterizes local frequency band information and global statistical information;

[0129] ;

[0130] Among them, Conv 3×3 () represents a 3×3 convolution (32 channels), BN() represents batch normalization, ReLU() represents the activation function, and GAP() represents global average pooling; thus, the resulting frequency band feature vector P k It is a channel description vector of length 32, which effectively extracts local and global statistical information of each frequency band.

[0131] S233 concatenates all frequency band feature vectors, predicts the weights of each spectral band through a fully connected layer, and then weights and fuses them to output differential features.

[0132] ;

[0133] in, Indicates the weights of the first fully connected layer; d represents the real number field; d represents the number of intermediate channels (e.g., 64). This represents the weights of the second fully connected layer; the Softmax function guarantees the weight vector... Satisfy: The weight of the k-th spectrum segment ;

[0134] Finally, the fused output features are calculated as follows:

[0135] ;

[0136] This process enables dynamic selection and fusion of different frequency band components, and supports Top-k sparse activation (such as taking only the two frequency bands with the highest weights), further reducing memory and computational load.

[0137] To prevent the network from relying solely on a few frequency band components for an extended period, this scheme uses weight balancing to assist in loss calculation. The loss function of the BandRouter dynamically weighted network is as follows:

[0138] ;

[0139] ;

[0140] in, This represents the loss function of the Band Router dynamically weighted network. This represents the average activation weight of the k-th spectral segment within a batch, where N represents the batch size. This represents the prediction weight of the nth sample in the kth spectral segment.

[0141] After determining the frequency bands, it is necessary to decide "which band is more important." This solution uses a Band Router dynamic weighted network to establish a lightweight 3×3 convolution-BN function-ReLU function branch for each frequency band, and calculates the frequency band feature (band tensor) F. k By applying 3×3 convolution, BN function, ReLU function, and GAP function, the frequency band feature vector P is obtained. k ; Put all frequency band feature vectors P k After concatenation, the input consists of two fully connected layers and a sigmoid function. This compresses the spatial information to a 1×1 shape, and the concatenation results in a vector. The weights are then predicted and output as a weight vector using the fully connected layers and the sigmoid function. These weights can adaptively change for different samples: high-detail scenes amplify high frequencies, while cloud or low-resolution scenes tend to retain low-to-mid frequencies. Finally, bilinear interpolation is used to calculate the weights according to block scale B. s Upsampled to the original size; after dimensionality reduction by 1×1 convolution, it is output as a frequency domain enhancement feature for subsequent spatial-frequency domain fusion.

[0142] All frequency band features are linearly fused according to the weight vector and resynthesized into a single tensor, maintaining 64C channels and a resolution of (H / 8, W / 8). The fused frequency domain features need to be restored to the same size as the backbone network. Bilinear interpolation is used to spatially enlarge the tensor by a factor of 8, restoring it to H×W. At this point, the channel dimension is still 64C, so a 1×1 convolution is then used to compress the channels back to their original C, completing dimension matching and adding a nonlinear mapping to help the frequency domain enhancement information blend naturally with the subsequent spatial convolutional stream. This scheme uses a lightweight router network to predict dynamic weights, automatically highlighting key signals and suppressing noisy frequency bands.

[0143] The Band Router in this application works in conjunction with the adaptive frequency band partitioning module. The former (frequency band partitioning) dynamically determines the coverage of each frequency band through a learnable threshold, while the latter (Band Router) adaptively adjusts the contribution of each frequency band feature based on the actual spectral distribution of the input image. Together, they form a dynamic expert system of "frequency division-frequency selection," capable of adaptively adjusting the optimal feature combination in various scenarios such as urban areas, farmland, and water bodies, enhancing the model's adaptability and discriminative ability to complex and changing images.

[0144] In summary, the AFB_DCT_FEM model provides an adjustable but orthogonal frequency domain dictionary based on a learnable 2D DCT transform network. The adaptive frequency band division AFB network softly segments the spectrum according to a trainable threshold. The Band Router dynamic weighting network scores each spectrum segment according to the sample content and weights and fuses them. Finally, it returns to the spatial domain through upsampling and 1×1 convolution. This approach retains the discriminative power brought by frequency analysis while avoiding the rigidity and gradient instability of fixed frequency band design.

[0145] Step 3: The preprocessed first and second temporal images are fed in parallel into the spatial encoder to extract multi-scale spatial features; in this embodiment, the spatial encoder is a conventional encoder, such as a ResNet encoder or a Swin-UNet encoder.

[0146] Step four involves fusing the differential features with the multi-scale spatial features to obtain a pixel-level variation mask. Specifically, in the multi-scale stage, strategies such as attention coupling, feature alignment, and differential accumulation are used to fuse the differential features and multi-scale spatial features, and finally, a lightweight decoder outputs the pixel-level variation mask.

[0147] The above network or model is trained using an end-to-end training method. During the training process, binary cross-entropy + DiceLoss is used as the main loss. At the same time, an orthogonal regularization term is introduced for the learnable DCT basis, and an ordered prior regularization is introduced for the adaptive threshold, so that the overall network automatically converges under the drive of the supervision signal.

[0148] The AFB_DCT_FEM model, located at the network front end, is responsible for "projecting the original spatial domain image onto a learnable frequency domain representation and extracting efficient differential frequency bands." Compared to existing methods that only perform differential operations on fixed high-frequency channels, the AFB_DCT_FEM model has three major innovations: First, it uses a depthwise separable convolutional DCT to simulate a learnable 2-D DCT, and the DCT convolution kernel can be adjusted iteratively during training; second, it uses a trainable scalar sequence to segment the frequency band domain, achieving soft frequency band division; and third, it designs lightweight CNN + global pooling prediction weights for the differential features of each frequency band, achieving scene-adaptive spectrum segment reconstruction.

[0149] Example 2

[0150] This embodiment provides a variation segmentation system based on learnable DCT and frequency band division, such as Figure 4 As shown, the system for implementing the variation segmentation method based on learnable DCT and frequency band division described in Example 1 includes:

[0151] The acquisition module is used to acquire first and second phase images of the same target area;

[0152] The preprocessing module is used to preprocess the first temporal image and the second temporal image;

[0153] The first feature extraction module is used to input the preprocessed first and second temporal images into the AFB_DCT_FEM module to obtain block-level DCT coefficient tensors, and generate differential features through adaptive frequency band division and routing weighting.

[0154] The second feature extraction module is used to send the preprocessed first temporal image and the second temporal image into the spatial encoder in parallel to extract multi-scale spatial features.

[0155] The fusion output module is used to fuse the differential features with the multi-scale spatial features to obtain a pixel-level variation mask.

[0156] Example 3

[0157] This embodiment provides a computer-readable medium storing a computer program that, when executed by a processor, can implement the variation segmentation method based on learnable DCT and frequency band division as described in Embodiment 1; specifically, it performs the following steps:

[0158] Step 1: Acquire a first temporal image and a second temporal image of the same target area, and preprocess the first temporal image and the second temporal image;

[0159] Step 2: Input the preprocessed first and second temporal images into the AFB_DCT_FEM model to obtain the block-level DCT coefficient tensor, and generate differential features through adaptive frequency band division and routing weighting.

[0160] Step 3: The preprocessed first and second temporal images are fed into the spatial encoder in parallel to extract multi-scale spatial features.

[0161] Step four: Fuse the differential features with the multi-scale spatial features to obtain a pixel-level variation mask.

[0162] Based on the proposed method, image change region segmentation is performed. The comparison results of the previous image, the later image, the ground truth of the change, and the model output are as follows: Figure 5 As shown.

[0163] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A segmentation method based on learnable DCT and frequency band division, characterized in that, The methods include: Acquire a first-phase image and a second-phase image of the same target area, and preprocess the first-phase image and the second-phase image; The preprocessed first and second temporal images are input into the AFB_DCT_FEM model to obtain the block-level DCT coefficient tensor, and differential features are generated through adaptive frequency band division and routing weighting. The preprocessed first and second temporal images are fed into the spatial encoder in parallel to extract multi-scale spatial features. The differential features are fused with the multi-scale spatial features to obtain a pixel-level variation film; The method for obtaining the differential features includes: Based on a learnable 2D DCT transform network, a cross-domain fine-tuning strategy is used to perform block-level frequency domain mapping on either the first or second temporal image, resulting in a mapped spectrum containing multiple frequency points; the DCT convolution kernel of the learnable 2D DCT transform network is a trainable DCT convolution kernel, and the DCT basis functions are: ； ； Where x and y represent the spatial coordinate indices within the block; B s The convolution block size is represented by u and v; the frequency domain coordinates are represented by u and v. The length weight of the k-th spectrum segment is indicated; k = u or v; cos[] represents the cosine function; The AFB network based on adaptive frequency band division performs frequency segmentation on the mapped spectrum according to the trainable frequency band threshold. The Band Router dynamic weighted network scores and weights each spectrum segment obtained from the segmentation, and then upsamples to obtain differential features.

2. The variation segmentation method based on learnable DCT and frequency band division according to claim 1, characterized in that, The preprocessing includes: The first and second temporal images are corrected, including radiometric correction and geometric correction. The corrected first and second temporal images are cropped to a target block size, which is a size divisible by a block of size B×B.

3. The variation segmentation method based on learnable DCT and frequency band division according to claim 1, characterized in that, The cross-domain fine-tuning strategy includes: Initialize the DCT convolution kernel, construct the adaptive attenuation coefficient matrix of the target image sensor, and scale the DCT convolution kernel based on the adaptive attenuation coefficient matrix after loading the pre-trained weights. The spatial resolution of the target image and the baseline spatial resolution during the pre-training stage of the learnable 2D DCT transform network are obtained. The size of the convolutional block is determined by combining the baseline spatial resolution and the spatial resolution of the target image to perform multiple interpolation scaling operations on the DCT convolutional kernel. The size of the convolutional block satisfies the following: ； in, Indicates the size of the scaled convolutional block; r base Indicates the reference spatial resolution; r target B represents the spatial resolution of the target image; s Indicates the size of the convolution block; The gain function for each frequency point is determined based on the dark channel prior method, and the DCT convolution kernel is scaled according to the frequency point based on the gain function.

4. The variation segmentation method based on learnable DCT and frequency band division according to claim 3, characterized in that, The method based on dark channel priors determines the gain function for each frequency point, and the DCT convolution kernel is scaled by frequency point based on the gain function, including the following methods: Atmospheric light transmittance t was estimated based on the dark channel prior method; The gain function for each frequency point is calculated based on the following formula: ； ； in, B represents the frequency radius, k represents the hyperparameter controlling the gain intensity; s Indicates the size of the convolution block; exp[*] represents the exponentiation operation; Represents the gain function; u and v represent frequency domain coordinates; Scaling the DCT convolution kernel by frequency point based on the gain function at each frequency point: ； in, This represents the final learnable DCT convolutional weights that take effect. This represents the learnable DCT convolution weights adapted after training.

5. The variation segmentation method based on learnable DCT and frequency band division according to claim 1, characterized in that, The adaptive frequency band segmentation (AFB) network segments the mapped spectrum according to a trainable frequency band threshold; including... method: Configure the number of spectrum segments K, and the set of edge thresholds for the spectrum segments is as follows: in, Let represent the threshold of the (K-1)th spectral segment, satisfying ; The unsorted learnable scalars are identified, and the length weights of each spectral segment are obtained by combining the softmax function. The ordered boundary thresholds are then obtained based on prefix accumulation. For each spectral band, a soft mask is generated based on the double sigmoid difference function.

6. The variation segmentation method based on learnable DCT and frequency band division according to claim 5, characterized in that, The Band Router-based dynamic weighted network scores and weights each segmented spectrum band, and finally upsamples to obtain differential features. Including methods: Obtain the frequency band characteristics of K frequency bands ; For the k-th frequency band feature F k The frequency band feature vector P is extracted through lightweight convolution. k The frequency band feature vector P k It characterizes local frequency band information and global statistical information; All frequency band feature vectors are concatenated, and the weights of each frequency band are predicted through a fully connected layer. The weighted fusion is then used to output differential features.

7. The variation segmentation method based on learnable DCT and frequency band division according to claim 6, characterized in that, The loss function of the Band Router dynamically weighted network is: ； ； in, This represents the loss function of the Band Router dynamically weighted network. This represents the average activation weight of the k-th spectral segment within a batch, where N represents the batch size. This represents the prediction weight of the nth sample in the kth spectral segment.

8. A segmentation system based on learnable DCT and frequency band division, characterized in that, The system is used to implement the variation segmentation method based on learnable DCT and frequency band division as described in any one of claims 1-7, the system comprising: The acquisition module is used to acquire first and second phase images of the same target area; The preprocessing module is used to preprocess the first temporal image and the second temporal image; The first feature extraction module is used to input the preprocessed first and second temporal images into the AFB_DCT_FEM module to obtain block-level DCT coefficient tensors, and generate differential features through adaptive frequency band division and routing weighting. The second feature extraction module is used to send the preprocessed first temporal image and the second temporal image into the spatial encoder in parallel to extract multi-scale spatial features. The fusion output module is used to fuse the differential features with the multi-scale spatial features to obtain a pixel-level variation film.

9. A computer-readable medium having a computer program stored thereon, characterized in that, The computer program, when executed by a processor, can implement the variation segmentation method based on learnable DCT and frequency band division as described in any one of claims 1-7.

Images