Remote sensing image mosaic dislocation detection method

Abstract

The invention relates to the technical field of remote sensing image processing, in particular to a remote sensing image mosaic dislocation detection method, which comprises the following steps: training: acquiring a data set and constructing a training sample set; pre-constructing a classification model based on the improved classification network, and training the classification model by using a two-stage training method and the training sample set until a trained target classification model is obtained; reasoning: obtaining a mosaic line and an orthoimage of a certain region of the to-be-detected image, extracting two side regions containing mosaic line features, and cutting according to a preset size to obtain a plurality of image blocks; and inputting all the image blocks as input images into the target classification model, and outputting a dislocation detection result used for describing whether dislocation exists in the to-be-detected image or not. According to the method, the image blocks are cut along the mosaic line, the image areas on the two sides of the mosaic line are extracted, whether dislocation exists or not is automatically judged through the improved classification model, and the final detection result is output in a vectorization mode and used for subsequent dislocation repairing or quality evaluation.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a remote sensing image mosaic misalignment detection method. Background Art

[0002] Traditional mosaic misalignment detection technologies primarily include gradient-based, corner-based, and spectral phase shift estimation-based methods. However, these three traditional methods suffer from insufficient detection accuracy, poor adaptability to complex scenes, and low efficiency, making them difficult to meet the practical needs of high-precision, large-scale remote sensing image mosaic misalignment detection.

[0003] To this end, some researchers have further introduced deep learning technology into mosaic misalignment detection, automatically extracting deep features from the images on both sides of the mosaic line and learning complex misalignment patterns, thereby achieving accurate and efficient detection. For example: 1) In "Blind Evaluation of Remote Sensing Image Mosaic Quality Based on Weak Supervision," Pan et al., published in the Journal of Remote Sensing, designed a dual-branch network to extract and learn features of color differences and ground structure misalignment on both sides of the mosaic seam, respectively. This approach employed a two-stage weakly supervised learning strategy to address the high cost of remote sensing image annotation. Furthermore, this method constructed a large-scale training set using simulated data and simultaneously calibrated the model with real annotated data, attempting to accurately evaluate complex mosaic problems with limited data. However, this method suffers from two drawbacks: first, the model's generalization ability is weak due to data issues; second, the model struggles to capture detailed features in complex scenes, which affects detection accuracy. 2) The Chinese patent, publication number CN119048520A, titled "Image Mosaic Misalignment Detection Method Based on Satellite Imagery and VGG Model Fine-tuning," focuses on detecting misalignment during satellite image mosaicking due to geometric errors, feature offsets, and discontinuous seams. This method uses VGG as a feature extraction network, but its feature extraction and representation capabilities are relatively limited. This makes it difficult to capture detailed information on both sides of the mosaicking line, particularly when processing high-resolution and complex features, impacting detection accuracy and robustness.

[0004] In addition to the aforementioned methods, some researchers have also adopted different technical approaches, such as deep learning, human collaboration, and optimization algorithms, to address issues such as color difference detection, mosaic editing, and mosaic line optimization in image mosaics. However, these methods are not suitable for the automated identification and detection of mosaic misalignment. For example, 1) Chinese patent publication number CN119360052A, entitled "Automated Detection Method for Color Difference in Large-Area Remote Sensing Image Mosaics," introduces a global response normalization (GRN) mechanism to enhance the model's global representation of image color difference features, enabling automatic and efficient detection and annotation of color difference regions on both sides of the mosaic line. However, this method is not suitable for mosaic misalignment detection tasks because mosaic misalignment is essentially a mismatch in the geometric position of the image. ConvNeXt, enhanced by the global response normalization mechanism, focuses on extracting global image color information and lacks sensitivity to local geometric misalignment and texture structure differences, making it difficult to effectively capture and identify geometric misalignment in images. 2) The Chinese patent with publication number CN118886868A, entitled "Orthophoto Collaborative Color-Uniform Mosaic Editing Method, System, Device and Medium," has achieved efficient manual detection and correction of color differences in the large-scale orthophoto mosaic process through reasonable task division, personnel management and collaborative operation. However, this technology relies entirely on manual visual discrimination and does not involve automatic algorithm detection. Therefore, it cannot be used for mosaic misalignment detection, which requires automatic identification of local geometric position misalignment. 3) The Chinese patent with publication number CN111311622A, entitled "Optimal Selection Method for Multiple Image Mosaic Line Network Based on Ant Colony Algorithm," automatically determines the optimal mosaic line position by analyzing the grayscale differences and texture features of the images to achieve seamless splicing between images. However, this technology does not involve the identification or detection of geometric misalignments that may occur during the mosaic process, and is therefore not suitable for mosaic misalignment detection tasks.

[0005] There are also some image stitching and defect detection methods in the industrial field, but each has certain limitations and is not suitable for automated detection of mosaic misalignment in remote sensing images. For example: 1) Chinese patent publication number CN114022450A, entitled "A stitching effect determination method for vehicle panoramic surround view testing," uses preset detection rules to identify gaps, misalignment, ghosting, and loss. However, this method only applies to well-defined rules in vehicle scenes and is not adaptable to the complex misalignment conditions found in remote sensing images. 2) Chinese patent publication number CN116309314A, entitled "Statistical-based vehicle panoramic surround view stitching image quality calculation method and system," uses checkerboard area segmentation and cluster fitting to detect misalignment. However, this method relies on manual placement of special checkerboard auxiliary equipment, limiting its versatility. 3) Chinese patent publication number CN116486178A, entitled "A defect detection method, device, electronic device, and storage medium," determines defect areas by analyzing the distance from pixel features to feature centers in the image to be inspected. This method is only applicable to industrial workpieces with relatively fixed structures and well-defined features. 4) The Chinese patent application, CN118691954A, titled "Automotive 360-degree Panoramic Image Stitching Quality Analysis Method Based on Machine Vision," analyzes changes in stitching quality using real-time driving speed. However, it primarily focuses on the stability of stitching quality with speed and is not suitable for detecting geometric misalignment in static remote sensing images.

[0006] In response to the above problems, those skilled in the art urgently need to propose a new mosaic misalignment detection method. Summary of the Invention

[0007] Therefore, the technical problem to be solved by the present invention is to overcome the defects existing in the above-mentioned prior art, thereby providing a remote sensing image mosaic misalignment detection method.

[0008] A remote sensing image mosaic misalignment detection method comprising: Training: Obtain a data set and construct a training sample set; pre-build a classification model based on the improved classification network, and train the classification model using a two-stage training method and a training sample set until a trained target classification model is obtained; Reasoning: Obtain the mosaic line and orthophoto of a certain area of ​​the image to be detected, extract the areas on both sides containing the mosaic line features, and crop them according to the preset size to obtain several image blocks; All image blocks are input into the target classification model as input images, and the output is a misalignment detection result used to describe whether there is misalignment in the image to be detected.

[0009] Preferably, the classification model is trained using a two-stage training method and a training sample set, specifically including: First stage training: Negative samples in the dataset are randomly sampled so that the ratio of positive and negative samples used in the first stage training is 1:1, and data augmentation methods are used for enhancement processing; the classification model is trained using the training sample set obtained from the processed dataset, and further optimized using the loss function and training optimizer to obtain the first model; Second stage training: After obtaining the model weights of the first model as the initialization weights, the classification model is retrained using all the positive and negative samples in the dataset. At the same time, the data augmentation method, loss function and training optimizer are also used in the training process, as well as the training sample set obtained based on the dataset to finally obtain the target classification model.

[0010] Preferably, the improved classification network is: using ConvNeXt as the basic network, further introducing a Fourier edge detection module for synchronous processing of the model input image and four pre-built edge pre-processing modules connected to the Fourier edge detection module; An edge attention module connected to the corresponding edge preprocessing module is added at the input of each of the four stages of the ConvNeXt backbone network.

[0011] Preferably, the network structure of the edge preprocessing module is convolution-batch normalization-activation-pooling connected in sequence.

[0012] Preferably, the network structure of the edge attention module includes two parts: the first part is edge feature-convolution-batch normalization-convolution-activation connected in sequence; The second part is the intermediate features; The intermediate feature is multiplied by the activation of the first part to obtain the product result, and the intermediate feature is then summed with the product result.

[0013] Preferably, obtaining a data set and constructing a training sample set specifically includes: Extract the dislocated image area and the restored non-dislocated image area from the pre-acquired orthophoto, crop them, annotate them, and save them into a pre-established empty set to obtain a dataset; Eliminate images in the dataset whose labels do not correspond to whether misalignment occurs; Randomly sample the negative samples in the dataset to control the ratio of the number of positive samples to the number of negative samples in the dataset; the images that are misaligned in the dataset are positive samples, and the images that are not misaligned are negative samples; The image sizes in the dataset are normalized, and images exceeding the size threshold are removed, ultimately obtaining the dataset required by the classification model.

[0014] Preferably, the negative samples in the data set are randomly sampled to control the ratio of the number of positive samples to the number of negative samples in the data set to 1:100.

[0015] The technical solution of the present invention has the following advantages: This paper uses ConvNeXt as the underlying architecture for its classification model, but its sensitivity to local geometric misalignment and texture differences in remote sensing imagery is insufficient. Therefore, the paper designs an edge attention module to enhance the classification model's sensitivity to texture details and edge features, significantly improving the accuracy of misalignment detection. Furthermore, a two-stage training method is designed to mitigate the imbalance between positive and negative samples. This method provides a highly efficient solution for mosaic quality control in DOM production, significantly enhancing the automation level and application value of large-scale remote sensing image data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 Schematic diagram of the overall network structure of the classification model of the present invention; Figure 2 Schematic diagram of important nodes in the training phase and inference phase in Example 2 of the present invention; Figure 3 This is an example diagram of the detection results of Example 2 of the present invention. DETAILED DESCRIPTION

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0019] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0020] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed, detachable, or integral connections; mechanical or electrical connections; direct or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0021] In addition, the technical features involved in different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0022] Example 1 Before specifically introducing the technical solution of this embodiment, the relevant technical concepts are introduced for easy understanding. Digital orthophoto map, also known as Digital Orthophoto Map, or DOM for short, is a remote sensing image product that has undergone geometric correction. It eliminates the influence of factors such as terrain undulations, sensor tilt, and perspective distortion, and accurately locates each pixel in the image to its actual position on the ground, ensuring that the image has a uniform scale and true geographic coordinates. This allows DOM to not only have rich image details, but also be directly used in geographic information systems, also known as Geographic Information System, or GIS for short, analysis and mapping. DOM is widely used in the fields of land and resources surveys, agricultural monitoring, urban planning, environmental protection, traffic management, and disaster emergency response, providing important spatial information support and having extremely high social and economic value.

[0023] Mosaicing technology is an essential step in the DOM production process. Its primary task is to stitch multiple remote sensing images together into a continuous, comprehensive image atlas based on geographic location. In large-scale DOM production, mosaicking can significantly improve image integration efficiency. However, due to errors in the geometric correction of remote sensing images, the influence of terrain undulations, and registration errors, image misalignment can easily occur on either side of the mosaicking line. This mosaicking misalignment not only compromises the geometric accuracy of the image but also affects the visual quality, thereby diminishing the application value of DOM and creating difficulties for users in data analysis and decision-making.

[0024] Traditional mosaic misalignment detection techniques primarily include gradient-based, corner-based, and spectral phase shift estimation-based methods. Gradient-based detection algorithms determine misalignment areas by calculating image gradient changes. However, in complex terrain scenes, gradient changes are easily affected by texture similarity and noise, resulting in inaccurate detection results. Corner-based detection algorithms rely on matching significant image features, such as corners, to determine misalignment. However, when the texture of the images on either side of the mosaic line is unclear or the terrain features are sparse, this method's stability and robustness are significantly reduced. Furthermore, spectral-based phase shift estimation algorithms estimate mosaic error by analyzing the frequency domain characteristics of the image, typically using changes in phase difference in the frequency domain to determine misalignment areas. However, this method is highly dependent on the integrity of the spectral information and the regularity of the frequency distribution. In situations with complex image textures or large terrain variations, the frequency domain features may lack clear periodicity, making it difficult to accurately estimate the phase difference, thus affecting detection accuracy. Furthermore, this method places high demands on the symmetry of the images on both sides of the mosaic line, making it difficult to adapt to scenes with large terrain undulations or uneven distribution of terrain features.

[0025] To address the above issues, this embodiment proposes a mosaic misalignment detection method. This method crops image blocks along the mosaic line, extracts image areas on both sides of the mosaic line, and uses an improved classification model to automatically determine whether there is misalignment. The final detection result is output in vectorized form for subsequent misalignment repair or quality assessment. A remote sensing image mosaic misalignment detection method includes: Training: Obtain a data set and construct a training sample set; pre-build a classification model based on the improved classification network, and train the classification model using a two-stage training method and a training sample set until a trained target classification model is obtained; Reasoning: Obtain the mosaic line and orthophoto of a certain area of ​​the image to be detected, extract the areas on both sides containing the mosaic line features, and crop them according to the preset size to obtain several image blocks; All image blocks are input into the target classification model as input images, and the output is a misalignment detection result used to describe whether there is misalignment in the image to be detected.

[0026] Specifically: Obtain the data set and construct a training sample set, including: Extract the dislocated image area and the restored non-dislocated image area from the pre-acquired orthophoto, crop them, annotate them, and save them into a pre-established empty set to obtain a dataset; Eliminate images in the dataset whose labels do not correspond to whether misalignment occurs; Negative samples in the dataset are randomly sampled to control the ratio of the number of positive samples to the number of negative samples in the dataset; among them, images that are misaligned in the dataset are positive samples; images that are not misaligned are negative samples.

[0027] The image sizes in the dataset are normalized, and images exceeding the size threshold are removed, ultimately obtaining the dataset required by the classification model.

[0028] Negative samples in the dataset are randomly sampled to control the ratio of the number of positive samples to the number of negative samples in the dataset to be 1:100.

[0029] The classification model is trained using a two-stage training method and a training sample set, specifically including: First stage training: Negative samples in the dataset are randomly sampled so that the ratio of positive and negative samples used in the first stage training is 1:1, and data augmentation methods are used for enhancement processing; the classification model is trained using the training sample set obtained from the processed dataset, and further optimized using the loss function and training optimizer to obtain the first model; Second stage training: After obtaining the model weights of the first model as the initialization weights, the classification model is retrained using all the positive and negative samples in the dataset. At the same time, the data augmentation method, loss function, and training optimizer are also used in the training process, as well as the training sample set obtained based on the dataset, and finally the target classification model is obtained.

[0030] The improved classification network is as follows: using ConvNeXt as the basic network, it further introduces a Fourier edge detection module that processes the model input image synchronously and four pre-built edge pre-processing modules connected to the Fourier edge detection module; An edge attention module connected to the corresponding edge preprocessing module is added at the input of each of the four stages of the ConvNeXt backbone network.

[0031] The network structure of the edge preprocessing module is convolution-batch normalization-activation-pooling connected in sequence.

[0032] The network structure of the edge attention module consists of two parts: the first part is edge feature-convolution-batch normalization-convolution-activation connected in sequence; The second part is the intermediate features; The intermediate feature is multiplied by the activation of the first part to obtain the product result, and the intermediate feature is then summed with the product result.

[0033] Specifically: The classification model network structure is as shown in the attached Figure 1As shown in the figure, the classification model takes an 8-bit remote sensing image of size 224×224×3 as input and progressively extracts features through a series of modules, ultimately determining whether there is any misalignment in the input image. First, these input images are passed through the Fourier edge detection module to extract high-frequency edge features. The Fourier edge detection module converts the input image into a grayscale image, then performs a two-dimensional fast Fourier transform (FFT) to obtain a frequency domain representation. A high-pass filter is then applied to mask low-frequency components, retaining only the image's edge information. Finally, an inverse Fourier transform is performed to restore the edge feature map. After normalization, a single-channel edge feature map with a size of [224, 224, 1] is output.

[0034] The first part of the backbone network is convolution and layer normalization. The input image is convolved with a kernel size of 4×4 and a step size of 4 to achieve downsampling, while adjusting the number of channels. This step changes the input image feature map from [224, 224, 3] to [56, 56, 96]. The edge feature map is aligned with the image feature map after the edge preprocessing module in the first stage. The width and height of the edge feature map are adjusted to the same as the input image feature map. Figure 1 The edge preprocessing module performs convolution and pooling operations on the edge feature map to further encode and enhance feature expression capabilities and achieve multi-scale matching. The image feature map and edge feature map are fused via the edge attention module. The edge attention module performs two convolution operations on the edge features to generate an attention map. The attention weights are element-wise multiplied by the image features and then added back to the original image feature map to enhance the feature expression of the edge region. The fused feature map still has a size of [56, 56, 96].

[0035] Each of the four stages of the backbone network consists of several ConvNeXt modules. The core of the ConvNeXt modules includes depthwise separable convolutions for extracting spatial features, channel-wise normalization to stabilize feature distribution, and point-wise convolutions to enhance global feature representation in the channel dimension. The input feature map size in the first stage is [56, 56, 96], which remains unchanged after processing through three ConvNeXt modules. Subsequently, the image feature map is downsampled via convolution with a stride of 2. In the next stage, the image feature map size is reduced to [28, 28, 192]. Simultaneously, the edge feature map is resized to the same width and height via the corresponding edge preprocessing module. After downsampling at each stage, the image and edge feature maps are fused via the edge attention module to enhance feature representation in edge regions. Similarly, the image feature map size in subsequent stages is reduced to [14, 14, 384] and [7, 7, 768], respectively. Through multiple convolutions, attention fusion, and downsampling operations in each stage, multi-scale features with rich semantic information are ultimately extracted. In the final stage, the classification model performs global average pooling (GAP) on the image feature map, reduces the dimension of the image feature map to [1,768], and outputs the classification result through the fully connected layer.

[0036] The ConvNeXt module includes depth convolution-layer normalization-convolution-activation-convolution-layer scaling-regularization connected in sequence; and the sum of the image feature map of the input depth convolution and the regularization output result in the ConvNeXt module is the final output of the ConvNeXt module.

[0037] Example 2 Based on Example 1, this example further discloses the following detailed technical content: The two-stage training method and inference method of classification are shown in the attached Figure 2 As shown in the figure, the training phase can be divided into five main nodes: orthophoto, mosaic line and manual quality inspection; manual review; dataset construction; data enhancement and two-stage model training; The inference phase can be divided into four main nodes: orthophoto and mosaic lines of a certain area; data segmentation; model determination of whether there is dislocation; dislocation vector generation; The specific two-stage model training steps are as follows: First, the data set needs to be constructed. Manual quality inspection vectors are used to determine the problematic areas in the orthophoto, and mosaic line vectors are used to stitch and process these areas to generate high-quality orthophotos. The image areas with dislocations and the image areas without dislocations after restoration are extracted from the orthophotos, and they are cropped and saved separately. For positive samples marked as dislocated, the label is set to 1; for negative samples marked as not dislocated, the label is set to 0. After the sample collection is completed, the image and label are manually checked to see if they match. Samples that do not meet the correspondence are removed from the data set. Since the number of negative samples is significantly greater than that of positive samples, the negative samples are randomly sampled to reduce a large number of redundant negative samples, and finally the positive-to-negative sample ratio of the data set is controlled at approximately 1:100. Subsequently, the image sizes in the data set are normalized. For images with a width or height less than 224, their short sides are resampled to 224 and their long sides are proportionally adjusted to the corresponding size; for images with a width or height between 500 and 1000, their long sides are resampled to 500 and their short sides are proportionally adjusted to the corresponding size; for images with a width or height greater than 1000, they are removed from the dataset.

[0038] Next, the first phase of training for the classification network was conducted. Because the dataset had a significant class imbalance with a positive-to-negative sample ratio of approximately 1:100, negative samples were randomly sampled again during the first phase of training to ensure a roughly 1:1 ratio of positive to negative samples. Furthermore, to improve the model's generalization capabilities, data augmentation methods were introduced, including random horizontal flipping and random scaling and cropping. AdamW was selected as the training optimizer, and the cross-entropy loss function was used.

[0039] Finally, the second phase of model training begins. After loading the model weights from the first phase of training as initial weights, the model is retrained using all positive and negative samples. During training, data augmentation methods such as random horizontal flipping and random scaling and cropping are still introduced to improve model robustness. The optimizer and loss function remain the same as in the first phase: the AdamW optimizer and the cross-entropy loss function.

[0040] This embodiment constructs a data set according to the above method. Specifically, 60,000 positive samples and 16 million negative samples were initially collected. After data normalization and negative sample resampling, a standardized data set containing 57,000 positive samples and 4.96 million negative samples was finally formed. In the first stage of training, the number of positive samples was maintained at 57,000. At the same time, negative samples were randomly sampled, and 72,000 negative samples were selected. The data set was divided into a training set and a test set at a ratio of 8:2. In the second stage of training, all positive and negative samples were used, and the training set and the test set were also divided at a ratio of 8:2. In order to improve the generalization ability of the model, data enhancement methods were used in both stages of training, including random horizontal flipping and random scaling and cropping. AdamW was selected as the training optimizer, the loss function was the cross entropy loss function, and the maximum number of training rounds was set to 100.

[0041] The reasoning steps are as follows: First, crop the mosaic line area. The input is the mosaic line and orthophoto of a region. The image is expanded to the left and right by a specified width, centering on the mosaic line. The regions containing the mosaic line features are extracted and cropped into 224×224 image blocks. To avoid missing critical information at the cropping boundary, a 20-30 pixel overlap is set between adjacent image blocks during cropping. The cropped image blocks are saved as tif files in preparation for classification model inference.

[0042] Then, inference and misalignment determination are performed, using all cropped image blocks as input to the trained classification model in batches. The classification model first uses a Fourier edge detection module to extract edge features for each image. Integrating this module with a dynamic edge attention module within the backbone network, the model extracts multi-scale image features layer by layer. The final classification layer of the classification model determines whether each image block exhibits mosaic misalignment. For misalignment determination, the classification model returns a binary classification result, indicating "misalignment present" or "no misalignment," and generates a corresponding judgment label for each image block.

[0043] Finally, the misaligned regions are merged and vectors are generated. After obtaining the inference results for all image blocks, the range of the image blocks marked as "misaligned" is extracted. The vectors of these misaligned regions are merged based on the geographic coordinate information of the image blocks to generate a complete vectorization result, representing all the misaligned areas in the survey area. Finally, the vectors of these misaligned regions are saved for subsequent mosaic line restoration or quality assessment.

[0044] To verify the effectiveness of the classification model designed in Example 1, several different models were compared under the same experimental setup. Table 1 shows the classification performance of different models on the test set, including the classification model designed in this example, the reference model ConvNeXt, and the ResNet-18 and VGG models used in related research.

[0045] Table 1 Classification performance of different models on the test set

[0046] To verify the effectiveness of the two-stage training strategy designed in this embodiment, the classification performance of the classification model under the two-stage training and one-stage training strategies is compared, as shown in Table 2.

[0047] Table 2 Classification performance of classification models with different training methods

[0048] In order to verify the effectiveness of the method proposed in this embodiment, it is applied in the quality inspection link of a certain orthophoto production project. The following are some examples of the detected misalignment: Figure 3 shown.

[0049] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will readily appreciate that other variations or modifications based on the above descriptions are possible. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A remote sensing image mosaic misalignment detection method, characterized in that: include: Training: Obtain the data set and construct a training sample set; Pre-build a classification model based on the improved classification network, and train the classification model using a two-stage training method and a training sample set until a trained target classification model is obtained; Reasoning: Obtain the mosaic line and orthophoto of a certain area of ​​the image to be detected, extract the areas on both sides containing the mosaic line features, and crop them according to the preset size to obtain several image blocks; All image blocks are input into the target classification model as input images, and the output is a misalignment detection result used to describe whether there is misalignment in the image to be detected.

2. The remote sensing image mosaic misalignment detection method according to claim 1, characterized in that: The classification model is trained using a two-stage training method and a training sample set, specifically including: First stage training: Negative samples in the dataset are randomly sampled so that the ratio of positive and negative samples used in the first stage training is 1:1, and data augmentation methods are used for enhancement processing; the classification model is trained using the training sample set obtained from the processed dataset, and further optimized using the loss function and training optimizer to obtain the first model; Second stage training: After obtaining the model weights of the first model as the initialization weights, the classification model is retrained using all the positive and negative samples in the dataset. At the same time, the data augmentation method, loss function and training optimizer are also used in the training process, as well as the training sample set obtained based on the dataset to finally obtain the target classification model.

3. The remote sensing image mosaic misalignment detection method according to claim 1, characterized in that: The improved classification network is as follows: using ConvNeXt as the basic network, it further introduces a Fourier edge detection module that processes the model input image synchronously and four pre-built edge pre-processing modules connected to the Fourier edge detection module; An edge attention module connected to the corresponding edge preprocessing module is added at the input of each of the four stages of the ConvNeXt backbone network.

4. A remote sensing image mosaic misalignment detection method according to claim 3, characterized in that: The network structure of the edge preprocessing module is convolution-batch normalization-activation-pooling connected in sequence.

5. The remote sensing image mosaic misalignment detection method according to claim 3, characterized in that: The network structure of the edge attention module consists of two parts: the first part is edge feature-convolution-batch normalization-convolution-activation connected in sequence; The second part is the intermediate features; The intermediate feature is multiplied by the activation of the first part to obtain the product result, and the intermediate feature is then summed with the product result.

6. The remote sensing image mosaic misalignment detection method according to claim 1, characterized in that: Obtain the data set and construct a training sample set, including: Extract the dislocated image area and the restored non-dislocated image area from the pre-acquired orthophoto, crop them, annotate them, and save them into a pre-established empty set to obtain a dataset; Eliminate images in the dataset whose labels do not correspond to whether misalignment occurs; Randomly sample the negative samples in the dataset to control the ratio of the number of positive samples to the number of negative samples in the dataset; the images that are misaligned in the dataset are positive samples, and the images that are not misaligned are negative samples; The image sizes in the dataset are normalized, and images exceeding the size threshold are removed, ultimately obtaining the dataset required by the classification model.

7. The remote sensing image mosaic misalignment detection method according to claim 6, characterized in that: Negative samples in the dataset are randomly sampled to control the ratio of the number of positive samples to the number of negative samples in the dataset to be 1:100.

Images