Remote sensing image intelligent quality detection method and system based on image segmentation

Abstract

The application relates to the technical field of image processing, and discloses a remote sensing image intelligent quality detection method and system based on image segmentation. The method comprises the following steps: acquiring a multispectral remote sensing image and performing adaptive region growing segmentation to obtain a ground object region block; a quality defect feature vector is extracted by improving a Vision Transformer network; a distortion defect is detected based on geometric consistency analysis; various quality problems are identified by using a multi-branch classification network; and a quality detection report is generated by reverse mapping and score calculation. The application establishes an accurate mapping mechanism from region-level detection to pixel-level positioning and a comprehensive quantitative evaluation system, thereby improving the accuracy, comprehensiveness and practicability of remote sensing image quality detection.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method and system for intelligent quality detection of remote sensing images based on image segmentation. Background Technology

[0002] Remote sensing images, as a crucial means of acquiring information about the Earth's surface, play a vital role in surveying, environmental monitoring, and urban planning. Traditional remote sensing image quality inspection methods mainly rely on manual visual interpretation or automated inspection based on statistical indicators. These methods are inefficient and highly subjective when processing large-scale remote sensing data. With the development of deep learning technology, image quality inspection methods based on convolutional neural networks have begun to be applied in the field of remote sensing. These methods train classifiers to identify quality defects in images, such as geometric distortion, radiation anomalies, and noise pollution, offering improvements in automation and detection accuracy compared to traditional methods.

[0003] However, existing intelligent quality inspection methods for remote sensing images based on image segmentation have significant technical shortcomings. These shortcomings mainly manifest as problems in detection results due to sample limitations, including three aspects: limited error types, imbalanced sample size, and poor scene adaptability. Limited error types mean that existing methods tend to overemphasize certain types of quality defects while failing to adequately identify other types. Imbalanced sample size leads to low accuracy in identifying some rare but important defect types. Poor scene adaptability causes a significant drop in detection performance when faced with remote sensing images acquired from different geographical environments and sensors. Furthermore, while they can detect some quality problems such as distortion, their accuracy and robustness remain insufficient when dealing with complex geometric deformations and subtle quality defects. Summary of the Invention

[0004] This application provides a remote sensing image intelligent quality inspection method and system based on image segmentation. It addresses the technical problems of limited sample types, imbalanced sample numbers, and poor scene adaptability in existing technologies by constructing a multi-scale semantic segmentation algorithm specifically for remote sensing image features, a quality defect-sensitive Vision Transformer network, and a multi-branch quality defect classification system. Furthermore, this application improves the accuracy, comprehensiveness, and practicality of remote sensing image quality inspection by establishing a precise mapping mechanism from region-level detection to pixel-level localization and a comprehensive quantitative evaluation system.

[0005] In a first aspect, this application provides a remote sensing image intelligent quality detection method based on image segmentation, the remote sensing image intelligent quality detection method based on image segmentation includes:

[0006] Step S1: Acquire multispectral remote sensing images, and perform semantic segmentation processing on the multispectral remote sensing images using an adaptive region growing segmentation algorithm based on the spectral features of ground objects to obtain a set of remote sensing ground object regions.

[0007] Step S2: Input the set of remote sensing ground object blocks into the improved Vision Transformer feature extraction network, and perform deep feature extraction on each remote sensing ground object block by introducing semantic location encoding and quality defect sensitive attention mechanism to obtain a remote sensing quality defect feature vector group.

[0008] Step S3: Based on the remote sensing quality defect feature vector group, calculate the principal direction angle deviation and boundary curvature value of adjacent remote sensing ground object blocks through the geometric consistency analysis algorithm to obtain the distortion deformation defect detection result;

[0009] Step S4: Input the remote sensing quality defect feature vector group and the distortion defect detection result into the multi-branch quality defect classification network, and perform classification processing for geometric distortion, radiation anomaly, noise pollution and data missing, respectively, to obtain the probability distribution of remote sensing image quality defects;

[0010] Step S5: Based on the probability distribution of quality defects in the remote sensing image, the quality defects are accurately located and quantitatively evaluated through reverse spatial mapping and comprehensive scoring calculation to obtain a remote sensing image quality inspection report.

[0011] Secondly, this application provides a remote sensing image intelligent quality detection system based on image segmentation, the remote sensing image intelligent quality detection system based on image segmentation includes:

[0012] The segmentation module is used to acquire multispectral remote sensing images and perform semantic segmentation processing on the multispectral remote sensing images using an adaptive region growing segmentation algorithm based on the spectral features of ground objects to obtain a set of remote sensing ground object regions.

[0013] The extraction module is used to input the set of remote sensing ground feature blocks into the improved Vision Transformer feature extraction network, and to perform deep feature extraction on each remote sensing ground feature block by introducing semantic location encoding and quality defect sensitive attention mechanism to obtain a remote sensing quality defect feature vector group.

[0014] The detection module is used to calculate the principal direction angle deviation and boundary curvature value of adjacent remote sensing ground object blocks based on the remote sensing quality defect feature vector group and through the geometric consistency analysis algorithm, so as to obtain the distortion deformation defect detection result;

[0015] The classification module is used to input the remote sensing quality defect feature vector group and the distortion defect detection results into a multi-branch quality defect classification network, and perform classification processing for geometric distortion, radiation anomaly, noise pollution and data missing, respectively, to obtain the probability distribution of remote sensing image quality defects.

[0016] The positioning module is used to accurately locate and quantitatively evaluate the quality defects based on the probability distribution of the remote sensing image quality defects through reverse spatial mapping and comprehensive scoring calculation, and to obtain a remote sensing image quality inspection report.

[0017] Thirdly, a remote sensing image intelligent quality detection device based on image segmentation is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the remote sensing image intelligent quality detection device based on image segmentation to execute the aforementioned remote sensing image intelligent quality detection method based on image segmentation.

[0018] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the above-described intelligent quality detection method for remote sensing images based on image segmentation.

[0019] The technical solution provided in this application utilizes an adaptive region growing segmentation algorithm based on the spectral features of ground features. This invention fully leverages the unique advantages of multispectral information in remote sensing images, intelligently segmenting based on the spectral differences of the red, green, blue, and near-infrared bands. Compared to traditional fixed-threshold segmentation methods, it can more accurately identify the boundaries of different ground feature types, avoiding the fragmentation of ground feature objects caused by blind segmentation, thus laying a more accurate spatial semantic foundation for subsequent quality detection. The improved Vision Transformer feature extraction network, by introducing semantic location coding and a quality defect-sensitive attention mechanism, overcomes the limitations of the local receptive field of traditional convolutional networks. It can establish long-distance dependencies between image patches. In particular, the semantic location coding generates specialized codes based on the spatial distribution characteristics of different ground feature types such as buildings, roads, and vegetation, enabling the network to learn prior knowledge related to ground features. The quality defect-sensitive attention mechanism enhances the attention to potential distortion and radiation anomaly areas through a priori weight matrix, significantly improving the targeting and effectiveness of feature learning. The geometric consistency analysis algorithm calculates geometric descriptors such as principal direction angle deviation and boundary curvature value, and is specifically designed to detect distortion and deformation of the geometric characteristics of regular features such as buildings in remote sensing images. Compared with traditional methods that rely solely on pixel-level statistical features, it has a stronger geometric semantic understanding ability and can accurately identify complex geometric distortion problems caused by factors such as satellite attitude changes and atmospheric turbulence.

[0020] The design of the multi-branch quality defect classification network fully demonstrates the professional advantages of this invention in the field of specific remote sensing image quality detection. By designing four specialized branches for geometric distortion, radiation anomaly, noise contamination, and data missing, each branch is optimized for the feature patterns of the corresponding quality defect type. The geometric distortion branch combines distortion detection results and deep features for classification, the radiation anomaly branch integrates RGB three-channel statistical features for color and contrast analysis, and the noise contamination branch extracts frequency domain features through discrete wavelet transform for noise identification. This multi-branch parallel processing architecture effectively solves the limitation of traditional single classifiers in handling multiple quality problems, and significantly improves the recognition accuracy and robustness of various defects. The introduction of reverse spatial mapping and comprehensive scoring calculation has achieved a technological breakthrough from regional detection to pixel-level precise positioning. By establishing a two-way coordinate mapping relationship, quality defects are accurately located to specific pixel positions in the original image. Combined with a comprehensive quantitative evaluation mechanism of pixel area ratio, severity coefficient, and confidence coefficient, it can not only identify the existence of quality problems, but also quantitatively assess their impact and spatial distribution. This provides objective and reliable technical support for the quality control and subsequent applications of remote sensing images. The overall solution shows significant technical advantages in solving key technical problems such as the single error type caused by sample limitations, the imbalance of sample quantity, and poor scene adaptability. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of an embodiment of the intelligent quality detection method for remote sensing images based on image segmentation in this application.

[0023] Figure 2 This is a schematic diagram of an embodiment of the intelligent quality detection system for remote sensing images based on image segmentation in this application.

[0024] Figure 3 This is a schematic block diagram of the structure of the remote sensing image intelligent quality detection device based on image segmentation in an embodiment of the present invention. Detailed Implementation

[0025] This application provides a method and system for intelligent quality inspection of remote sensing images based on image segmentation. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent quality detection method for remote sensing images based on image segmentation in this application includes:

[0027] Step S1: Acquire multispectral remote sensing images, and perform semantic segmentation on the multispectral remote sensing images using an adaptive region growing segmentation algorithm based on the spectral features of ground objects to obtain a set of remote sensing ground object regions.

[0028] Step S2: Input the set of remote sensing ground object blocks into the improved Vision Transformer feature extraction network, and perform deep feature extraction on each remote sensing ground object block by introducing semantic location encoding and quality defect sensitive attention mechanism to obtain a remote sensing quality defect feature vector group.

[0029] Step S3: Based on the remote sensing quality defect feature vector group, calculate the main direction angle deviation and boundary curvature value of adjacent remote sensing ground object blocks through the geometric consistency analysis algorithm to obtain the distortion deformation defect detection results;

[0030] Step S4: Input the remote sensing quality defect feature vector group and the distortion and deformation defect detection results into the multi-branch quality defect classification network, and perform classification processing for geometric distortion, radiation anomaly, noise pollution and data missing, respectively, to obtain the probability distribution of remote sensing image quality defects;

[0031] Step S5: Based on the probability distribution of quality defects in remote sensing images, the quality defects are accurately located and quantitatively evaluated through reverse spatial mapping and comprehensive scoring calculation to obtain a remote sensing image quality inspection report.

[0032] It is understood that the executing entity of this application can be a remote sensing image intelligent quality inspection system based on image segmentation, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiments use a server as an example for illustration.

[0033] Specifically, spectral feature vectors are extracted from the input multispectral remote sensing image. Each pixel's spectral feature vector contains numerical information from four bands: red, green, blue, and near-infrared. The algorithm measures spectral similarity by calculating the Euclidean distance between adjacent pixels. When the difference in the spectral feature vectors of two adjacent pixels is less than a preset threshold, these two pixels are grouped into the same growth region. The region growth process starts from the seed pixel and gradually expands outwards until it encounters a boundary with a significant spectral difference. Subsequently, semantic label matching is performed on the generated initial segmented regions. Based on the spectral feature patterns of different land cover types, each segmented region is labeled as one of the following: building, road, vegetation, water body, or bare soil, forming a semantic label set. Finally, the spatial coordinate information of each remote sensing land cover region is recorded, establishing a correspondence between the region block and the pixel location in the original image.

[0034] Deep feature learning of remotely sensed land cover regions is performed using an improved Vision Transformer feature extraction network. First, each region is standardized to a uniform size of 224×224 pixels, and then divided into 196 image patches of 16×16 pixels each. These image patch sequences undergo positional encoding, including standard positional encoding and semantic positional encoding generated based on land cover type. Semantic positional encoding can distinguish the spatial distribution characteristics of different land cover types in remotely sensed images, enabling the network to learn prior knowledge related to land cover types. The encoded image patch sequences are input into the improved Vision Transformer network, where the feature dimension of each image patch is transformed to 768 dimensions through a linear projection layer. The network constructs a query matrix, a key matrix, and a value matrix, and introduces a prior weight matrix specifically designed for remotely sensed image quality defects. This weight matrix is ​​constructed based on the statistical characteristics of common quality problems in remotely sensed images, such as distortion and radiation anomalies, enhancing the attention given to areas with potential quality defects during attention calculation. Through iterative calculations using a 12-layer transformer encoder, each layer incorporates a multi-head self-attention mechanism and a feedforward neural network, ultimately yielding a 768-dimensional feature vector rich in quality information.

[0035] A geometric consistency analysis algorithm is employed to specifically detect distortion defects. Based on the obtained remote sensing quality defect feature vector set, a local geometric descriptor is extracted for each remote sensing feature block, including three key parameters: principal orientation angle, aspect ratio, and boundary curvature. For building-type blocks, the algorithm analyzes whether their principal orientation angle conforms to the geometric rules of the building. When the detected angle deviates from the standard angle values ​​of 0, 45, 90, or 135 degrees by more than 8 degrees, the area is determined to have an angle distortion problem. The aspect ratio detection is achieved by analyzing the length relationship of adjacent parallel edges. Under normal circumstances, the lengths of opposite edges of a building should be equal. When the length ratio exceeds the normal range of 0.95 to 1.05, it is marked as length distortion. Boundary curvature detection identifies geometric deformation by calculating the curvature of straight line segments. When the curvature value exceeds the threshold of 0.03, it is determined to be bending deformation. The algorithm integrates and analyzes the three types of detection results—angle distortion, length distortion, and bending deformation—to form a complete distortion defect detection result.

[0036] A multi-branch quality defect classification network is constructed, with specialized processing branches designed for different types of quality problems. The geometric distortion classification branch receives remote sensing quality defect feature vectors and distortion / deformation defect detection results, and performs accurate classification of geometric distortion types through a three-layer fully connected network, outputting probability distributions including angular distortion, length distortion, bending deformation, and normal state. The radiation anomaly classification branch extracts statistical features of the RGB three channels from remote sensing ground feature blocks, calculates the mean, variance, and skewness parameters of each channel, and fuses these radiation features with deep features for radiation quality analysis, identifying problems such as color shift and contrast anomalies. The noise pollution classification branch performs frequency domain analysis on remote sensing ground feature blocks using discrete wavelet transform, extracting high-frequency coefficients as noise features, and combining them with deep features to identify noise types such as salt-and-pepper noise and Gaussian noise. The data missing classification branch specifically handles data integrity issues such as cloud cover obstruction and missing sensor stripes. The probability vectors output by each branch are weighted and fused using preset weight coefficients to obtain a comprehensive probability distribution of remote sensing image quality defects.

[0037] Precise location and quantitative assessment are performed based on the probability distribution of quality defects. A bidirectional mapping relationship is established, linking the coordinate system of remotely sensed ground feature blocks with the pixel coordinate system of the original remotely sensed image, enabling detected quality defects to be accurately mapped back to their specific pixel locations in the original image. The reverse spatial mapping process calculates the accurate boundary contours of defects in the original image based on the spatial coordinate information of the blocks, forming pixel-level defect annotations. The quantitative assessment module calculates the pixel area ratio, severity coefficient, and detection confidence of each quality defect region, obtaining an objective quality score through a comprehensive scoring formula. Finally, the quality scores are divided into different levels, generating a detection report that includes a statistical histogram of defect types, a spatial distribution heatmap, and a quality level assessment.

[0038] In one specific embodiment, step S1 includes:

[0039] Spectral feature vector extraction is performed on multispectral remote sensing images to obtain spectral feature vectors for each pixel containing values ​​of four bands: red, green, blue, and near-infrared.

[0040] The Euclidean distance difference between adjacent pixels is calculated based on the spectral feature vector. When the difference is less than a preset threshold, the adjacent pixels are grouped into the same region for region growing to obtain the initial segmentation region.

[0041] The initial segmented regions are processed by semantic tag matching, and each segmented region is labeled as one of the following categories: buildings, roads, vegetation, water bodies, and bare soil, to obtain a set of semantic tags;

[0042] Based on the semantic tag set, the initial segmented region is processed by spatial coordinate recording to obtain a set of remote sensing ground feature blocks and their corresponding spatial coordinate information.

[0043] Specifically, the spectral feature vector of each pixel contains values ​​for four bands: red, green, blue, and near-infrared. The red, green, and blue bands reflect the visible light reflectance characteristics of ground features, while the near-infrared band reflects information such as chlorophyll absorption by vegetation and soil moisture content. The construction process of the spectral feature vector first reads the location of each pixel in the original remote sensing image, extracts the corresponding values ​​for the four bands, and then combines these four values ​​to form a four-dimensional vector. The red band values ​​typically vary between 0 and 255, reflecting the intensity of red light reflection by ground features; the green and blue bands also use the same value range. The near-infrared band's value range is consistent with the visible light band, but its physical meaning differs, primarily used to distinguish between vegetated and non-vegetated areas. The spectral feature vector formed by the combination of the four band values ​​becomes the input data for subsequent region growing algorithms, with each pixel corresponding to a unique four-dimensional spectral signature.

[0044] The similarity of the spectral features of adjacent pixels determines whether pixels are classified into the same growth region. The Euclidean distance calculation process takes the spectral feature vectors of two adjacent pixels, calculates the differences in the four band values ​​for each, and then takes the square root of the sum of the squares of these four differences to obtain the final distance. When the spectral feature vector of pixel A is 120 for the red band, 85 for the green band, 70 for the blue band, and 150 for the near-infrared band, and the corresponding values ​​of adjacent pixel B are 115 for red, 88 for green, 72 for blue, and 145 for near-infrared, the Euclidean distance is equal to the square root of the sum of the squares of the red difference (5), the green difference (3), the blue difference (2), and the near-infrared difference (5). The preset threshold is set based on statistical analysis of the spectral features of different land cover types. Spectral differences in building areas are usually small, while spectral differences in vegetation areas are relatively large due to variations in chlorophyll content. When the calculated Euclidean distance is less than a preset threshold, the algorithm groups adjacent pixels into the same growth region, expands outward from the seed pixel until it encounters a boundary pixel whose spectral difference exceeds the threshold, and forms an initial segmentation region with similar spectral characteristics.

[0045] The semantic label matching process associates initial segmented regions with predefined land cover types, addressing the inaccurate land cover category identification problem in traditional methods. The matching process is based on typical spectral feature patterns of different land cover types. Building areas typically exhibit relatively balanced values ​​across the red, green, and blue bands with low near-infrared reflectance. Road areas have similar spectral characteristics to buildings but generally lower reflectance. Vegetation areas are characterized by significant red band absorption and strong near-infrared reflectance. Water bodies show strong absorption across all bands, while bare soil areas exhibit stronger red band reflectance than green and blue band reflectance. The algorithm calculates the average spectral feature vector of all pixels within each initial segmented region to obtain a representative spectral signature for that region, then compares it with the standard spectral patterns of the five land cover types. The similarity comparison uses a spectral angle matching method, calculating the angle between the region's average spectral vector and the standard spectral pattern. The land cover type with the smallest angle is identified as the semantic label for that region. The semantic label set contains the land cover type identifier corresponding to each segmented region, providing prior knowledge for subsequent quality detection algorithms. Spatial coordinate recording establishes a precise correspondence between segmented regions and their spatial locations in the original image, ensuring that quality detection results can accurately pinpoint specific geographical locations. The recording process traverses each initial segmented region, extracting the row and column coordinates of all pixels constituting that region, and calculating the region's bounding box, centroid coordinates, and geometric parameters. The bounding box is determined by finding the minimum and maximum row and column coordinates of the pixels within the region. The centroid coordinates are obtained by arithmetically averaging the coordinates of all pixels within the region. Geometric parameters include descriptive statistics such as region area, perimeter, and aspect ratio. Spatial coordinate information is stored in the form of structured data. Each remote sensing feature block corresponds to a complete set of spatial description parameters, including region identifier, feature type label, bounding box coordinates, centroid location, area size, and shape features. This spatial coordinate information plays a crucial role in subsequent Vision Transformer feature extraction and quality defect localization, ensuring that the algorithm can accurately identify and locate various quality problems in complex remote sensing images.

[0046] In one specific embodiment, step S2 includes:

[0047] Each block in the remote sensing ground feature block set is size-standardized, and all blocks are uniformly adjusted to a 224×224 pixel specification to obtain standardized blocks.

[0048] The standardized region blocks are divided into image blocks according to the 16×16 pixel specification. Each standardized region block generates 196 image blocks, resulting in an image block sequence.

[0049] Based on spatial coordinate information, the image patch sequence is overlaid with standard location coding and semantic location coding. The semantic location coding is generated according to the corresponding land cover type in the semantic label set to obtain the coded image patch sequence.

[0050] The encoded image patch sequence is input into the improved Vision Transformer feature extraction network. The feature learning process is performed through a multi-head self-attention layer containing a quality defect prior weight matrix. After iterative calculation by a 12-layer transformer encoder, the remote sensing quality defect feature vector group is obtained.

[0051] Specifically, this paper addresses the issue of inconsistent sizes in remote sensing feature patches by standardizing them. The standardization process analyzes the original size of each patch, including the width and height in pixels, and then uses bilinear interpolation to adjust all patches to a uniform 224×224 pixel size. The bilinear interpolation algorithm determines the value of a new pixel by calculating the weighted average of the four nearest neighbor pixels surrounding the target pixel, with the weights inversely proportional to the distance between the target pixel and its four neighbors. When the original patch size is 300×180 pixels, the algorithm first calculates the horizontal and vertical scaling ratios: the horizontal scaling ratio is 224 divided by 300, and the vertical scaling ratio is 224 divided by 180. Then, interpolation is performed on the pixel value at each target location. The interpolation process preserves the spectral characteristics of the original image, ensuring that the red, green, blue, and near-infrared band values ​​of the features are not significantly distorted during the size transformation. The standardized patch size meets the strict requirements of the VisionTransformer network for input data format while maintaining the integrity of the spectral characteristics of different feature types.

[0052] Image patch segmentation converts standardized region blocks into a serialized data format that the Vision Transformer network can process, overcoming the limitation of traditional convolutional networks in capturing long-range dependencies. The segmentation process divides each 224×224 pixel standardized region block into non-overlapping smaller blocks according to a fixed 16×16 pixel size. Specifically, starting from the top left corner, each 16×16 pixel square region is extracted as an independent image block. Since 224 divided by 16 equals 14, each standardized region block generates 14 image blocks horizontally and 14 horizontally, for a total of 196 image blocks. Each image block contains 16 x 16 x 4 pixel values, where 4 represents the four bands: red, green, blue, and near-infrared, forming a 1024-dimensional original feature vector. The image block sequence is numbered and arranged from left to right and top to bottom to ensure that spatial location information is preserved during serialization. The image patch segmentation process converts two-dimensional image data into one-dimensional sequential data, enabling the Vision Transformer network to learn the feature relationships between different spatial locations through a self-attention mechanism. Location coding overlay processing addresses the inherent problem of insufficient location information in Vision Transformer networks by introducing standard location coding and semantic location coding, particularly considering the spatial distribution characteristics of different land cover types in remote sensing images. Standard location coding is generated using sine and cosine functions, assigning a unique location identifier to each location in the image patch sequence, with the coding dimension consistent with the subsequent feature vector dimension. Semantic location coding generates specific coding vectors based on the corresponding land cover type in the semantic label set obtained in step S1. Semantic coding for building types emphasizes geometric regularity, for vegetation types it highlights seasonal variations, for roads it focuses on linear continuity, for water bodies it reflects spectral absorption characteristics, and for bare soil types it reflects surface roughness. The generation of semantic location codes is based on statistical analysis of a large number of remote sensing image samples, establishing coding patterns by learning the spatial distribution regularities of different land cover types. The coding overlay process adds the standard location coding vector element-wise to the corresponding semantic location coding vector for each land cover type, resulting in a coded image patch sequence that integrates spatial location information and semantic information.

[0053] The Vision Transformer feature extraction network learns deep feature representations of quality problems in remote sensing images through an improved multi-head self-attention mechanism and a quality defect prior weight matrix. The network first transforms the 1024-dimensional original features of each image patch into a unified 768-dimensional feature representation using a linear projection layer. The projection process employs fully connected layers to achieve dimensionality transformation and feature mapping. The multi-head self-attention layer constructs a query matrix, a key matrix, and a value matrix. The query matrix extracts query information from the input features through linear transformation, the key matrix extracts key information for matching, and the value matrix contains the actual feature content. The quality defect prior weight matrix is ​​constructed based on the statistical characteristics of quality problems such as distortion and radiation anomalies in remote sensing images. The weight values ​​in the matrix reflect the correlation strength between quality defects occurring at different spatial locations. The attention weight calculation process involves matrix multiplication of the transposes of the query matrix and the key matrix, then adding the quality defect prior weight matrix, and normalizing using a softmax function to obtain the final attention weight distribution. The iterative computation process of the 12-layer transformer encoder deepens the feature representation layer by layer. Each layer includes multi-head self-attention computation and feedforward neural network processing. The stability of gradient propagation is ensured through residual connections and layer normalization.

[0054] In one specific embodiment, the encoded image patch sequence is input into an improved Vision Transformer feature extraction network. Feature learning is performed through a multi-head self-attention layer containing a quality defect prior weight matrix. After iterative computation by a 12-layer transformer encoder, a remote sensing quality defect feature vector set is obtained, including:

[0055] The encoded image patch sequence is processed by a linear projection layer to transform the dimension of each image patch into 768 dimensions, thus obtaining the projection feature matrix.

[0056] The query matrix, key matrix, and value matrix are constructed based on the projection feature matrix. The prior weight matrix of quality defects, which is constructed based on the statistical characteristics of distortion and radiation anomalies in remote sensing images, is introduced to perform attention weight calculation and processing to obtain the quality defect sensitive attention matrix.

[0057] The quality defect sensitive attention matrix and the value matrix are processed by matrix multiplication, and the residual connection calculation of the multi-head attention mechanism and the feedforward neural network is performed to obtain the output features of the single-layer encoder.

[0058] The output features of a single-layer encoder are processed by 12 iterations of encoder layer calculation. Each iteration includes multi-head self-attention calculation and feedforward network processing to obtain a set of remote sensing quality defect feature vectors.

[0059] Specifically, the linear projection layer's dimensionality transformation process converts the original features of the encoded image patch sequence into the standard feature representation format of the VisionTransformer network, resolving the technical problem of input data dimensionality mismatch. The dimensionality transformation process receives the original 1024-dimensional feature vector for each image patch, which is obtained by flattening 16×16×4 pixel data, where 4 represents the four bands: red, green, blue, and near-infrared. The linear projection layer employs a fully connected neural network structure with a weight matrix size of 1024×768 and a bias vector length of 768. Matrix multiplication maps the input 1024-dimensional feature vector to a 768-dimensional standard feature representation. During the transformation, each input feature is element-wise multiplied with the corresponding row of the weight matrix, then summed and the bias term is added to obtain one dimension component of the output feature. This process is repeated 768 times to obtain the complete output vector. The dimension of the projected feature matrix is ​​196×768, where 196 corresponds to the number of image patches generated by each standardized region block, and 768 is the feature dimension of each image patch. The projection process not only unifies the feature dimensions, but also enhances the feature representation capabilities related to remote sensing image quality detection tasks through the learning process of weight parameters, especially the sensitivity to geometric regularity and spectral consistency.

[0060] The construction of the query matrix, key matrix, and value matrix is ​​based on a triple linear transformation of the projected feature matrix, extracting different types of information for attention calculation. The query matrix is ​​obtained by multiplying the projected feature matrix by the query weight matrix, which has a size of 768×768. Its learned parameters can extract query vectors from the input features for searching relevant information. The key matrix is ​​generated by multiplying the projected feature matrix by the key weight matrix, which also has a 768×768 dimension and is responsible for extracting key feature information for matching queries. The value matrix is ​​generated by multiplying the projected feature matrix by the value weight matrix, which also has a 768×768 dimension and contains the actual feature content information. The quality defect prior weight matrix is ​​the core innovation of this invention. This matrix is ​​constructed based on statistical analysis of distortion and radiation anomalies in a large number of remote sensing image samples. It has a size of 196×196, and each element represents the correlation strength of quality defects occurring between corresponding spatial locations. The attention weight calculation process first involves matrix multiplication of the transposes of the query matrix and the key matrix to obtain the original attention score matrix. Then, the prior weight matrix for quality defects is added element-wise to the original attention score matrix to enhance the attention to potential quality defect areas. Finally, the matrix is ​​normalized using the softmax function to obtain the quality defect sensitive attention matrix.

[0061] The matrix multiplication operation between the quality defect sensitive attention matrix and the value matrix fuses attention weights and feature content to generate a feature representation containing quality defect sensitive information. During matrix multiplication, each row of the attention matrix represents the attention weight distribution of one image patch to all other image patches, and each row of the value matrix contains the feature content of the corresponding image patch. The result of multiplying the two matrices is a new feature representation for each image patch that incorporates global context information. The multi-head attention mechanism captures different types of feature relationships by computing multiple attention heads in parallel. Each attention head uses different query, key, and value weight matrices to focus on different aspects of quality issues such as geometric structure, spectral features, and texture information. The feedforward neural network contains two linear transformation layers and an activation function. The first layer maps 768-dimensional features to a higher-dimensional intermediate representation, and the activation function uses ReLU for non-linear transformation. The second layer maps the intermediate features back to the 768-dimensional output. Residual connections solve the gradient vanishing problem in deep networks by element-wise adding the input features to the network output features, ensuring effective transfer of feature information between network layers. The output features of a single-layer encoder integrate the global feature relationships of multi-head attention and the local feature transformations of the feedforward network, forming a deep understanding of the quality status of remote sensing images.

[0062] The 12-layer iterative encoder computation deepens the level of feature abstraction layer by layer, gradually learning semantic-level quality defect patterns from pixel-level features at the bottom layer to semantic-level features at the top layer. Each iteration includes complete multi-head self-attention computation and feedforward network processing, with residual connections and layer normalization ensuring training stability between iterations. The first layer encoder mainly learns the basic spatial relationships and spectral similarities between image patches. The intermediate layer encoder gradually abstracts geometric features such as ground object boundaries and shape regularities, while the deep layer encoder focuses on identifying complex quality defect patterns, such as distortion at building edges and radiation anomalies in vegetation areas. Each layer's computation updates the quality defect-sensitive attention matrix, enabling the network to understand the feature representations of different types of quality problems at a deeper level. The final remote sensing quality defect feature vector set contains 768-dimensional feature representations of each image patch obtained through 12 layers of deep learning. These feature vectors not only retain the spectral and spatial information of the original image but also incorporate specialized knowledge for remote sensing image quality detection tasks.

[0063] In one specific embodiment, step S3 includes:

[0064] Based on the remote sensing quality defect feature vector group, the local geometric descriptor of each remote sensing land cover block is extracted, including the main direction angle, aspect ratio and boundary curvature parameters, to obtain a set of geometric feature descriptions;

[0065] Based on the main direction angle parameters in the geometric feature description set, the angle deviation of the building type area block is calculated. When the detected angle deviates from the standard angle value by more than 8 degrees, it is marked as angle distortion, and the angle distortion detection result is obtained.

[0066] Based on the aspect ratio parameters in the geometric feature description set, geometric distortion judgment is performed on the ratio of the lengths of adjacent parallel sides. When the ratio exceeds the range of 0.95 to 1.05, it is marked as length distortion, and the length distortion detection result is obtained.

[0067] The boundary curvature parameters in the geometric feature description set are compared and analyzed with the preset straight line segment curvature threshold. When the curvature value is greater than 0.03, it is determined to be bending deformation. Combined with the angle twist detection results and length distortion detection results, the twist deformation defect detection results are obtained.

[0068] Specifically, potential shape anomalies are identified by analyzing the spatial geometric characteristics of each remotely sensed feature block. The extraction process for the principal direction angle parameter first detects edge pixels within the block, using the Sobel operator to calculate the gradient direction of each pixel. Then, the histogram distribution of the gradient directions of all edge pixels is statistically analyzed, and the peak value of the histogram corresponds to the principal edge direction angle of the block. For building blocks, there should normally be two main peaks, corresponding to the horizontal and vertical edges of the building, respectively. Geometric regularity is determined by analyzing the deviation of these two peak angles from the standard 90-degree angle. The aspect ratio parameter is obtained by analyzing the minimum bounding rectangle of the block. The algorithm first extracts the boundary contour of the block, then calculates the minimum rectangle that can completely contain the contour, and records the lengths of the long and short sides of the rectangle; the ratio of these two lengths is the aspect ratio parameter. Boundary curvature parameters are calculated by analyzing the curvature of the boundary lines of the region blocks. The algorithm calculates the local curvature value point by point along the boundary contour, using the three-point method to calculate the curvature of each boundary point. That is, the curvature is measured by the rate of change of the angle formed by the current point and the two adjacent points. The average curvature value of the entire boundary is used as the boundary curvature parameter of the region block. The geometric feature description set contains three geometric parameters: principal direction angle, aspect ratio, and boundary curvature, corresponding to all remotely sensed ground feature regions.

[0069] Angle deviation calculation and processing is specifically designed for geometric regularity checks of building-type area blocks. It identifies angle distortion issues by comparing the deviation of the actually detected principal direction angle with the theoretical standard angle. The theoretical standard angles for buildings include four directions: 0 degrees, 45 degrees, 90 degrees, and 135 degrees, corresponding to the horizontal edge, diagonal direction, and vertical edge of the building, respectively. The angle deviation calculation process first extracts the principal direction angle parameters of the building area block from the geometric feature description set, and then calculates the absolute difference between this angle and the nearest standard angle. When the principal direction angle of the building area block is 97 degrees, the deviation from the nearest standard angle of 90 degrees is 7 degrees. Since the deviation value is less than the set threshold of 8 degrees, it is determined that there is no angle distortion in this area. When the detected principal direction angle is 102 degrees, the deviation from the 90-degree standard angle is 12 degrees, exceeding the 8-degree threshold, and is marked as an abnormal angle distortion. The angle distortion detection results record the angle deviation status of each building area block, including four levels: normal, slight distortion, moderate distortion, and severe distortion. The greater the deviation, the more severe the geometric deformation of the building.

[0070] The length ratio geometric distortion detection process identifies building scale deformation by analyzing the length relationships of adjacent parallel edges. The process first identifies parallel edge pairs within a building block, extracting key line segments using edge detection and line fitting algorithms. Then, it determines which line segments form parallel relationships based on the similarity of their slopes. For rectangular buildings, opposite edges should have equal lengths. The algorithm calculates the length ratio of each pair of parallel edges, which should ideally be close to 1.0. A length ratio between 0.95 and 1.05 is considered normal; values ​​outside this range are marked as length distortion. A ratio less than 0.95 indicates shrinkage relative to the opposite edge, while a ratio greater than 1.05 indicates stretching relative to the opposite edge. The length distortion detection result includes the ratio information and distortion type identifier for all parallel edge pairs within each building block. Shrinkage and stretching distortions correspond to different geometric deformation modes.

[0071] Boundary curvature comparison analysis identifies bending deformation defects by comparing the actual measured boundary curvature values ​​with a preset straight line segment curvature threshold. The theoretical curvature value of a straight line segment should be 0, representing a perfectly straight boundary line. However, in actual remote sensing images, due to various influencing factors during the imaging process, even the edges of straight buildings may exhibit slight curvature fluctuations. The preset curvature threshold of 0.03 is determined based on statistical analysis of a large number of normal remote sensing image samples, representing the upper limit of curvature variation for normal straight line boundaries. When the boundary curvature parameter of a region exceeds the 0.03 threshold, it indicates that the straight boundary of that region has significant bending deformation, and is judged as a bending deformation anomaly. The severity of bending deformation is directly proportional to the magnitude of the curvature value; the larger the curvature, the more severe the bending. The final distortion deformation defect detection result is obtained by comprehensively considering the angular distortion detection result, the length distortion detection result, and the bending deformation judgment result. A logical OR operation is used to merge the three anomaly types; if any geometric anomaly exceeds the threshold, the region is marked as having a distortion deformation defect.

[0072] In one specific embodiment, step S4 includes:

[0073] The remote sensing quality defect feature vector group is input into the geometric distortion classification branch, and combined with the distortion and deformation defect detection results, the geometric distortion type is classified through a three-layer fully connected network to obtain the geometric distortion probability vector.

[0074] Based on the remote sensing quality defect feature vector group, RGB three-channel statistical features are extracted. The mean, variance and skewness features are fused with deep features and then input into the radiation anomaly classification branch for color shift and contrast anomaly classification processing to obtain the radiation anomaly probability vector.

[0075] Discrete wavelet transform is performed on the set of remotely sensed ground feature blocks to extract high-frequency coefficient noise features. The noise features are then fused with the remote sensing quality defect feature vector set and input into the noise pollution classification branch for noise type identification to obtain the noise pollution probability vector.

[0076] Based on preset weighting coefficients, the probability vectors of geometric distortion, radiation anomaly, noise contamination, and missing data are weighted and fused to obtain the probability distribution of remote sensing image quality defects.

[0077] Specifically, the remote sensing quality defect feature vector set is fused with the distortion and deformation defect detection results, and a three-layer fully connected network is used to achieve accurate classification of geometric distortion types. The branch first receives a 768-dimensional quality defect feature vector corresponding to each remote sensing feature area block. This vector contains deep semantic features learned by the Vision Transformer network. Simultaneously, it receives the distortion and deformation defect detection results generated in step S3, which include detection states of three geometric anomalies: angular distortion, length distortion, and bending deformation. The feature fusion process concatenates the 768-dimensional deep feature vector with the 3-dimensional geometric detection result vector to form a 771-dimensional comprehensive input feature. The three-layer fully connected network is designed with the first layer containing 256 neurons, the second layer containing 64 neurons, and the third layer containing 4 neurons, corresponding to the four geometric distortion types: angular distortion, length distortion, bending deformation, and normal state, respectively. During network training, each neuron in each layer undergoes a linear transformation with the input features through a weight matrix, then a non-linear mapping is performed using the ReLU activation function, and finally, the output is converted into a probability distribution form using the softmax function. The four components of the geometric distortion probability vector represent the probability values ​​of the region block belonging to various geometric distortion types. The sum of the probability values ​​equals 1, and the type corresponding to the highest probability is the geometric distortion determination result of the region block.

[0078] The RGB three-channel statistical feature extraction process calculates the color distribution characteristics of each region based on the original remote sensing feature block set, overcoming the limitations of traditional methods in identifying radiometric quality issues. The statistical feature extraction process calculates the mean, variance, and skewness parameters for the red, green, and blue color channels respectively. The mean reflects the average brightness level of the region, the variance describes the dispersion of the color distribution, and the skewness measures the asymmetry of the color distribution. For each region block, the red channel mean is obtained by calculating the arithmetic mean of the red components of all pixels within the region; the variance is obtained by dividing the sum of the squares of the differences between the red value of each pixel and the mean by the total number of pixels; and the skewness is calculated by dividing the third central moment by the cube of the standard deviation. The green and blue channels are calculated using the same method. A 9-dimensional statistical feature vector contains the mean, variance, and skewness parameters for each of the RGB channels. This vector is concatenated with a 768-dimensional remote sensing quality defect feature vector to form a 777-dimensional radiometric analysis input feature. The radiation anomaly classification branch processes the fused features using a fully connected neural network. The network structure consists of a 512-dimensional hidden layer, a 128-dimensional intermediate layer, and a 3-dimensional output layer. The three neurons in the output layer correspond to color shift, contrast anomaly, and normal states, respectively. The radiation anomaly probability vector records the probability distribution of the region patch under the three radiation states. The color shift probability reflects the degree of deviation between the region's color and the standard ground feature color, while the contrast anomaly probability represents the degree of abnormality in color contrast within the region.

[0079] Discrete wavelet transform noise feature extraction identifies high-frequency noise components in remotely sensed ground object blocks through frequency domain analysis. Daubechies wavelets are used to decompose each block into multiple scales. The wavelet transform process decomposes the original block into a low-frequency approximate sub-band and three high-frequency detail sub-bands. The horizontal detail sub-band captures horizontal edge information, the vertical detail sub-band captures vertical edge information, and the diagonal detail sub-band captures diagonal texture information. High-frequency coefficient noise feature extraction focuses on the energy distribution of the three high-frequency detail sub-bands. A 9-dimensional noise feature vector is formed by calculating the mean absolute value, standard deviation, and peak amplitude of the coefficients in each sub-band. The high-frequency energy distribution in normal image areas exhibits a regular pattern, while noise-contaminated areas show abnormal energy concentration or dispersion in certain frequency bands. Salt-and-pepper noise manifests as isolated extreme points in the high-frequency detail sub-bands, while Gaussian noise shows a uniform increase in energy across all frequency bands. Different types of noise have significant differences in their frequency domain characteristics. After the noise features are fused with the remote sensing quality defect feature vector, they are input into the noise pollution classification branch. This branch adopts a classifier structure specifically designed for noise patterns and outputs the probability distribution of four states: salt-and-pepper noise, Gaussian noise, strip noise, and no noise. The noise pollution probability vector reflects the degree to which the area block is affected by various noises.

[0080] The weighted fusion calculation process linearly combines the output probability vectors of the four classification branches according to preset weight coefficients to obtain a comprehensive remote sensing image quality defect probability distribution. The weight coefficients are set based on the importance analysis of different types of quality defects to the subsequent application of remote sensing images. Geometric distortion has the greatest impact on measurement accuracy and therefore has the highest weight; radiometric anomalies have the next highest weight due to their impact on visual interpretation quality; noise pollution has a smaller weight due to its impact on image sharpness; and data missing has a moderate weight due to its impact on information integrity. The weighted fusion process first multiplies the probability vectors of geometric distortion, radiometric anomalies, noise pollution, and data missing by their respective weight coefficients. Then, the weighted probability vectors are summed element-wise to obtain the final quality defect probability distribution. This probability distribution includes all possible quality defect types and their corresponding probabilities of occurrence. A higher probability value indicates a greater likelihood of the presence of that type of defect. A probability threshold is set to determine whether a certain area is marked as having a quality problem.

[0081] In one specific embodiment, step S5 includes:

[0082] Based on spatial coordinate information, a two-way mapping relationship is established from the block coordinates of remotely sensed ground features to the pixel coordinates of the original remotely sensed image. The defect locations in the probability distribution of quality defects in the remotely sensed image are accurately located to the original image coordinate system through reverse spatial mapping processing, resulting in pixel-level defect boundary contours.

[0083] The pixel area ratio, severity coefficient and confidence coefficient of each quality defect region are calculated based on the pixel-level defect boundary contour. The quantitative evaluation is then performed using a comprehensive scoring formula to obtain a quantitative score for the quality defect.

[0084] The quality defect quantitative scoring is divided into four levels: excellent, good, qualified, and unqualified. Detailed analysis data, including defect type statistics and spatial distribution heatmaps, is generated to obtain a remote sensing image quality inspection report.

[0085] Specifically, a precise correspondence between the coordinate system of remotely sensed ground feature regions and the pixel coordinate system of the original remotely sensed image is constructed based on spatial coordinate information to solve the technical problem of inaccurate location of quality defects. The mapping process first extracts the bounding box coordinates of each remotely sensed ground feature region in the original image, including the top-left and bottom-right pixel coordinates, region width, and height parameters. These parameters record the precise spatial position of the region in the original image. The forward mapping relationship is established by calculating the linear transformation relationship between the relative coordinates within the region and the absolute coordinates of the original image. The transformation parameters include the horizontal scaling factor, the vertical scaling factor, and the coordinate offset. The reverse mapping relationship is achieved by inversely calculating the transformation from the pixel coordinates of the original image to the relative coordinates of the region. The transformation process uses an affine transformation matrix to describe the correspondence between the coordinate systems. The reverse spatial mapping process converts the relative position information of each defect region in the probability distribution of quality defects in the remotely sensed image obtained in step S4 into absolute pixel coordinates in the original image. The mapping process processes each region marked as having quality defects in the probability distribution, calculating the precise pixel coordinates of the defects based on their relative position within the region and their absolute position in the original image. Pixel-level defect boundary contour generation forms a closed boundary polygon by connecting adjacent defect pixels. Contour extraction uses a boundary tracking algorithm to search pixel by pixel along the outer edge of the defect region and record the coordinate sequence of all pixels that constitute the complete boundary.

[0086] The quality defect quantitative assessment process calculates multiple key evaluation indicators based on pixel-level defect boundary contours, achieving objective quantification of quality defects through a comprehensive scoring mechanism. Pixel area proportion calculation involves counting the number of pixels contained within the defect boundary contour and dividing this number by the total number of pixels in the corresponding remote sensing feature area to obtain the area proportion parameter, reflecting the spatial extent of the defect within that area. The severity coefficient is determined based on the defect type and characteristic parameters. The severity of geometric distortion is calculated using a weighted combination of angle deviation, length distortion ratio, and curvature value; the severity of radiation anomalies is assessed through color shift amplitude and contrast variation; and the severity of noise pollution is measured by signal-to-noise ratio and noise distribution density. The confidence coefficient reflects the reliability of the detection results, calculated based on the probability value output by the classification network and the feature consistency index. Higher probability values ​​and better feature consistency result in a higher confidence coefficient. The comprehensive scoring formula weights and synthesizes the pixel area proportion, severity coefficient, and confidence coefficient. The weight allocation is determined based on the importance of different evaluation indicators to the final quality judgment: the area proportion weight reflects the defect's impact range, the severity weight reflects the degree of defect harm, and the confidence weight reflects the detection accuracy. The quality defect quantification score is obtained by mapping the weighted composite result to a standard score range of 0 to 100. The lower the score, the more serious the quality defect, and the higher the score, the better the quality status.

[0087] The quality grading process quantifies quality defects and scores them according to preset threshold ranges, generating standardized quality evaluation results and detailed analysis reports. The grading uses a four-level evaluation system: Excellent (scoring 90 points or above, indicating good quality with no obvious defects); Good (scoring 80-89 points, indicating minor quality issues that do not affect normal use); Acceptable (scoring 70-79 points, indicating moderate quality defects requiring attention); and Unacceptable (scoring below 70 points, indicating serious quality problems requiring rework). Defect type statistics are generated by summarizing the quality inspection results of all regions, including the number of geometrically distorted areas, radiation anomaly areas, noise pollution areas, and missing data areas. Each defect type is further subdivided into three sub-levels: mild, moderate, and severe. The spatial distribution heatmap visualizes the quality scores according to the spatial location of the original image, using color coding to represent the quality status of different areas: red areas indicate serious quality problems, yellow areas indicate average quality, and green areas indicate good quality. The remote sensing image quality inspection report integrates the quality grade classification results, defect type statistics, spatial distribution heat map, and detailed defect location list to form a complete quality assessment document. The report format adopts a standardized template to ensure the consistency and comparability of information expression.

[0088] The above describes the intelligent quality detection method for remote sensing images based on image segmentation in the embodiments of this application. The following describes the intelligent quality detection system for remote sensing images based on image segmentation in the embodiments of this application. Please refer to [link / reference]. Figure 2 One embodiment of the remote sensing image intelligent quality detection system based on image segmentation in this application includes:

[0089] The segmentation module is used to acquire multispectral remote sensing images and perform semantic segmentation processing on the multispectral remote sensing images using an adaptive region growing segmentation algorithm based on the spectral features of ground objects to obtain a set of remote sensing ground object regions.

[0090] The extraction module is used to input the set of remote sensing ground feature blocks into the improved Vision Transformer feature extraction network, and to perform deep feature extraction on each remote sensing ground feature block by introducing semantic location encoding and quality defect sensitive attention mechanism to obtain a remote sensing quality defect feature vector group.

[0091] The detection module is used to calculate the principal direction angle deviation and boundary curvature value of adjacent remote sensing ground object blocks based on the remote sensing quality defect feature vector group and through the geometric consistency analysis algorithm, so as to obtain the distortion deformation defect detection result;

[0092] The classification module is used to input the remote sensing quality defect feature vector group and the distortion defect detection results into a multi-branch quality defect classification network, and perform classification processing for geometric distortion, radiation anomaly, noise pollution and data missing, respectively, to obtain the probability distribution of remote sensing image quality defects.

[0093] The positioning module is used to accurately locate and quantitatively evaluate the quality defects based on the probability distribution of the remote sensing image quality defects through reverse spatial mapping and comprehensive scoring calculation, and to obtain a remote sensing image quality inspection report.

[0094] above Figure 2 The image segmentation-based intelligent quality inspection system for remote sensing images in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The image segmentation-based intelligent quality inspection device in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0095] Reference Figure 3 This invention also provides an intelligent quality inspection device for remote sensing images based on image segmentation. This device can be a server, and its internal structure can be as follows: Figure 3As shown, the image segmentation-based intelligent remote sensing image quality inspection device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor, designed as a computer, provides computational and control capabilities. The memory of the image segmentation-based intelligent remote sensing image quality inspection device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the image segmentation-based intelligent remote sensing image quality inspection device stores the data corresponding to this embodiment. The network interface of the image segmentation-based intelligent remote sensing image quality inspection device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0096] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the image segmentation-based remote sensing image intelligent quality detection device to which the present invention is applied.

[0097] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the image segmentation-based remote sensing image intelligent quality detection method.

[0098] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0099] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a remote sensing image intelligent quality inspection device based on image segmentation (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0100] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A remote sensing image intelligent quality detection method based on image segmentation, characterized in that, The method includes: Step S1: Acquire multispectral remote sensing images, and perform semantic segmentation processing on the multispectral remote sensing images using an adaptive region growing segmentation algorithm based on the spectral features of ground objects to obtain a set of remote sensing ground object regions. Step S2: Input the set of remotely sensed land cover regions into the improved Vision Transformer feature extraction network. By introducing semantic location encoding and a quality defect sensitive attention mechanism, deep feature extraction is performed on each remotely sensed land cover region to obtain a set of remotely sensed quality defect feature vectors. This includes: standardizing the size of each region in the set of remotely sensed land cover regions, adjusting each region to a uniform 224×224 pixel size to obtain standardized region blocks; dividing the standardized region blocks into image blocks according to a 16×16 pixel size, generating 196 image blocks for each standardized region block to obtain an image block sequence; performing standard location encoding and semantic location encoding superposition processing on the image block sequence based on spatial coordinate information, wherein the semantic location encoding is generated according to the corresponding land cover type in the semantic label set to obtain an encoded image block sequence; inputting the encoded image block sequence into the improved Vision Transformer feature extraction network, performing feature learning processing through a multi-head self-attention layer containing a quality defect prior weight matrix, and iteratively calculating through a 12-layer transformer encoder to obtain the set of remotely sensed quality defect feature vectors. Step S3: Based on the remote sensing quality defect feature vector group, calculate the principal direction angle deviation and boundary curvature value of adjacent remote sensing ground object blocks through the geometric consistency analysis algorithm to obtain the distortion deformation defect detection result; Step S4: Input the remote sensing quality defect feature vector group and the distortion defect detection result into the multi-branch quality defect classification network, and perform classification processing for geometric distortion, radiation anomaly, noise pollution and data missing, respectively, to obtain the probability distribution of remote sensing image quality defects; Step S5: Based on the probability distribution of quality defects in the remote sensing image, the quality defects are accurately located and quantitatively evaluated through reverse spatial mapping and comprehensive scoring calculation to obtain a remote sensing image quality inspection report.

2. The intelligent quality detection method for remote sensing images based on image segmentation according to claim 1, characterized in that, Step S1 includes: The multispectral remote sensing image is processed by spectral feature vector extraction to obtain a spectral feature vector containing values ​​of four bands: red, green, blue, and near-infrared for each pixel. Based on the spectral feature vector, the Euclidean distance difference between adjacent pixels is calculated. When the difference value is less than a preset threshold, adjacent pixels are grouped into the same region for region growing to obtain the initial segmentation region. The initial segmented regions are subjected to semantic tag matching processing, and each segmented region is labeled as one of the following: buildings, roads, vegetation, water bodies, and bare soil, to obtain a set of semantic tags; Based on the semantic tag set, the initial segmented region is processed to record spatial coordinates, thereby obtaining the remote sensing feature block set and its corresponding spatial coordinate information.

3. The intelligent quality detection method for remote sensing images based on image segmentation according to claim 2, characterized in that, The encoded image patch sequence is input into the improved Vision Transformer feature extraction network. Feature learning is performed through a multi-head self-attention layer containing a quality defect prior weight matrix. After iterative computation by a 12-layer transformer encoder, the remote sensing quality defect feature vector set is obtained, including: The encoded image block sequence is processed by a linear projection layer to transform the feature dimension of each image block into 768 dimensions, thus obtaining the projection feature matrix. Based on the projection feature matrix, a query matrix, a key matrix, and a value matrix are constructed. Then, a quality defect prior weight matrix constructed based on the statistical characteristics of remote sensing image distortion and radiation anomaly is introduced to perform attention weight calculation processing to obtain a quality defect sensitive attention matrix. The quality defect sensitive attention matrix and the value matrix are multiplied by matrix multiplication, and the residual connection calculation of the multi-head attention mechanism and the feedforward neural network is performed to obtain the output features of the single-layer encoder. The output features of the single-layer encoder are processed by 12 iterations of transformer encoder layer calculation. Each iteration includes multi-head self-attention calculation and feedforward network processing to obtain the remote sensing quality defect feature vector group.

4. The intelligent quality detection method for remote sensing images based on image segmentation according to claim 1, characterized in that, Step S3 includes: Based on the remote sensing quality defect feature vector group, local geometric descriptors of each remote sensing land cover area are extracted, including principal direction angle, aspect ratio and boundary curvature parameters, to obtain a set of geometric feature descriptions; Based on the main direction angle parameters in the geometric feature description set, the angle deviation of the building type area block is calculated. When the detected angle deviates from the standard angle value by more than 8 degrees, it is marked as angle distortion, and the angle distortion detection result is obtained. Based on the aspect ratio parameters in the geometric feature description set, the ratio of the lengths of adjacent parallel sides is subjected to geometric distortion judgment processing. When the ratio exceeds the range of 0.95 to 1.05, it is marked as length distortion, and the length distortion detection result is obtained. The boundary curvature parameters in the geometric feature description set are compared and analyzed with the preset straight line segment curvature threshold. When the curvature value is greater than 0.03, it is determined to be bending deformation. The result of the tortuous deformation defect detection is obtained by combining the angle distortion detection result and the length distortion detection result.

5. The intelligent quality detection method for remote sensing images based on image segmentation according to claim 1, characterized in that, Step S4 includes: The remote sensing quality defect feature vector group is input into the geometric distortion classification branch, and combined with the distortion and deformation defect detection results, geometric distortion type classification is performed through a three-layer fully connected network to obtain the geometric distortion probability vector. Based on the remote sensing quality defect feature vector group, RGB three-channel statistical features are extracted. The mean, variance and skewness features are fused with deep features and then input into the radiation anomaly classification branch for color shift and contrast anomaly classification processing to obtain the radiation anomaly probability vector. Discrete wavelet transform is performed on the remote sensing ground object block set to extract high-frequency coefficient noise features. The noise features are then fused with the remote sensing quality defect feature vector set and input into the noise pollution classification branch for noise type identification to obtain a noise pollution probability vector. The probability distribution of remote sensing image quality defects is obtained by weighting and fusing the geometric distortion probability vector, the radiation anomaly probability vector, the noise contamination probability vector, and the data missing probability vector according to preset weighting coefficients.

6. The intelligent quality detection method for remote sensing images based on image segmentation according to claim 3, characterized in that, Step S5 includes: Based on the spatial coordinate information, a two-way mapping relationship is established from the coordinates of remotely sensed ground features to the pixel coordinates of the original remotely sensed image. The defect locations in the probability distribution of quality defects in the remotely sensed image are accurately located to the original image coordinate system through reverse spatial mapping processing to obtain pixel-level defect boundary contours. Based on the pixel-level defect boundary contour, the pixel area ratio, severity coefficient and confidence coefficient of each quality defect region are calculated, and the quantitative evaluation is performed through a comprehensive scoring formula to obtain a quantitative score for the quality defect. The quality defect quantitative score is divided into four levels: excellent, good, qualified, and unqualified. Detailed analysis data, including defect type statistics and spatial distribution heatmaps, is generated to obtain the remote sensing image quality inspection report.

7. A remote sensing image intelligent quality inspection system based on image segmentation, characterized in that, For implementing the image segmentation-based intelligent quality detection method for remote sensing images as described in any one of claims 1 to 6, the image segmentation-based intelligent quality detection system for remote sensing images comprises: The segmentation module is used to acquire multispectral remote sensing images and perform semantic segmentation processing on the multispectral remote sensing images using an adaptive region growing segmentation algorithm based on the spectral features of ground objects to obtain a set of remote sensing ground object regions. The extraction module is used to input the set of remotely sensed ground object regions into an improved VisionTransformer feature extraction network. By introducing semantic location encoding and a quality defect-sensitive attention mechanism, deep feature extraction is performed on each remotely sensed ground object region to obtain a set of remotely sensed quality defect feature vectors. This includes: standardizing the size of each region in the set of remotely sensed ground object regions, adjusting each region to a uniform 224×224 pixel size to obtain standardized region blocks; dividing the standardized region blocks into image blocks according to a 16×16 pixel size, generating 196 image blocks per standardized region block to obtain an image block sequence; performing standard location encoding and semantic location encoding superposition processing on the image block sequence based on the spatial coordinate information, wherein the semantic location encoding is generated according to the corresponding ground object type in the semantic label set to obtain an encoded image block sequence; inputting the encoded image block sequence into the improved VisionTransformer feature extraction network, performing feature learning processing through a multi-head self-attention layer containing a quality defect prior weight matrix, and iteratively calculating through a 12-layer transformer encoder to obtain the set of remotely sensed quality defect feature vectors. The detection module is used to calculate the principal direction angle deviation and boundary curvature value of adjacent remote sensing ground object blocks based on the remote sensing quality defect feature vector group and through the geometric consistency analysis algorithm, so as to obtain the distortion deformation defect detection result; The classification module is used to input the remote sensing quality defect feature vector group and the distortion defect detection results into a multi-branch quality defect classification network, and perform classification processing for geometric distortion, radiation anomaly, noise pollution and data missing, respectively, to obtain the probability distribution of remote sensing image quality defects. The positioning module is used to accurately locate and quantitatively evaluate the quality defects based on the probability distribution of the remote sensing image quality defects through reverse spatial mapping and comprehensive scoring calculation, and to obtain a remote sensing image quality inspection report.

8. A remote sensing image intelligent quality inspection device based on image segmentation, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the intelligent quality detection method for remote sensing images based on image segmentation as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it causes the processor to execute the remote sensing image intelligent quality detection method based on image segmentation as described in any one of claims 1 to 6.

Images