Digital ray image splicing and identification method and system based on high-contrast characteristic

Abstract

The invention provides a digital ray image splicing and identification method and system based on high contrast characteristics. The method comprises the following steps: carrying out Gaussian filtering denoising and adaptive histogram equalization preprocessing on an aerospace digital ray image; extracting multi-scale high-contrast features by using a local binary pattern algorithm and carrying out series fusion; matching feature points through a scale invariant feature transformation algorithm, and estimating an affine transformation model in combination with a random sampling consistency algorithm to realize panoramic stitching; and constructing a convolutional neural network, and training the model by using a cross entropy loss function to complete panoramic image intelligent identification. The system correspondingly comprises a preprocessing module, a feature extraction module, a panoramic stitching module, an intelligent identification module and a data storage module. According to the method, the problems of low splicing precision and low recognition accuracy when the spaceflight digital ray image is processed by a traditional method are effectively solved, the image recognizability, the splicing precision and the target recognition accuracy are improved, and the requirements of detection of the internal quality of spaceflight parts and defect recognition are met.

Description

Technical Field

[0001] This invention relates to the field of aerospace nondestructive testing and image processing technology, specifically to a method and system for digital ray image stitching and recognition based on high-contrast features. More particularly, it relates to an efficient stitching and intelligent recognition method for aerospace digital ray panoramic images based on high-contrast features, which can be used for the detection of internal quality and defect identification of aerospace components. Background Technology

[0002] In the aerospace field, aerospace components are characterized by high structural complexity and large size ranges. Even minute internal defects (such as cracks, pores, and inclusions) directly impact the safety and reliability of aerospace missions. Digital radiographic (DR) technology, with its non-contact and high-penetration advantages, has become a core method for acquiring internal quality information of components. However, due to the limited coverage of a single inspection, multiple local digital radiographic images must be stitched together to create a panoramic view to achieve comprehensive inspection of large and complex components. Yet, aerospace digital radiographic images inherently suffer from low contrast, strong noise interference, and blurred details, posing significant challenges to the accuracy of image stitching and defect identification.

[0003] Current aerospace digital X-ray image processing and detection technology has formed a core technical framework of preprocessing-stitching-recognition, and related patents and research have explored various aspects of each stage, but significant technical limitations still exist. Regarding image preprocessing and feature extraction, patent application CN113884515A discloses an integrated control system for digital X-ray detection. Its image processing module only uses basic contrast adjustment and image correction methods, without optimizing for the low contrast and high noise characteristics of aerospace DR images. This makes it difficult to effectively highlight the detailed features of internal defects, and it lacks coordination with subsequent stitching and recognition stages. Traditional feature extraction algorithms such as SIFT and ORB are prone to low feature recognition accuracy due to blurred feature points and noise interference when processing such images. Related research literature also confirms that directly applying traditional feature extraction methods significantly increases the mismatch rate in subsequent matching stages. Regarding image stitching, the paper "Research on Image Stitching Technology Based on Deep Neural Networks" proposes a registration scheme based on a cross-attention matching module. While this improves the registration effect in special scenarios, it is not specifically adapted to the characteristics of digital ray images, does not combine high-contrast features to optimize matching accuracy, and does not solve the problem of adaptability of the stitched image for defect identification. Some existing stitching technologies only focus on the geometric alignment of images, neglecting the grayscale consistency and detail preservation of the stitched area, resulting in ghosting, misalignment, or loss of detail in the panoramic image, affecting the subsequent recognition effect. In terms of defect identification, existing technologies are mostly based on single local images for identification, with a limited field of view, making it difficult to capture the complete shape of internal defects in components. Even when using deep learning models, the insufficient feature quality of the input image (low contrast, blurred details) leads to low accuracy in identifying small defects, failing to meet the high-precision inspection requirements of the aerospace field.

[0004] From the perspective of technical implementation effectiveness, the core problems of existing technologies can be summarized into four points: First, preprocessing and feature extraction are not targeted enough. Traditional methods are difficult to balance denoising and detail preservation in aerospace DR images. High-contrast feature extraction lacks effective technical means, resulting in low differentiation between defects and background in the image, which poses hidden dangers for subsequent stitching and recognition. Second, the image stitching accuracy and stability are poor. Traditional feature matching algorithms are prone to mismatches in low-contrast, high-noise images. Deep neural network stitching methods are not adapted to the characteristics of DR images, resulting in problems such as geometric misalignment and uneven grayscale in the panoramic image, which cannot provide a high-quality global observation basis. Third, the global vision and feature utilization for defect recognition are insufficient. Existing recognition technologies mostly rely on local images and do not make full use of the global information of the stitched panoramic image. Furthermore, they do not combine high-contrast features to mine defect details, resulting in low recognition accuracy for small defects and complex-shaped defects. Fourth, there is a lack of full-process collaboration. Existing technologies separate preprocessing, stitching, and recognition into independent stages. The processing results of each stage cannot provide accurate support for subsequent stages. There is a lack of an end-to-end collaborative optimization framework, resulting in poor overall processing performance.

[0005] In summary, existing digital X-ray inspection image processing technologies cannot meet the high-precision and high-accuracy non-destructive testing requirements of aerospace components. There is an urgent need in the market for a full-process collaborative optimization method based on high-contrast features. This method would integrate targeted preprocessing, high-precision stitching, and intelligent recognition technologies to solve the problems of stitching accuracy and recognition accuracy of low-contrast and high-noise DR images, providing reliable technical support for the internal quality inspection of aerospace components. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a digital ray image stitching and recognition method and system based on high contrast features, so as to solve the problems of low stitching accuracy and low recognition accuracy of aerospace digital ray images in existing technologies.

[0007] A digital ray image stitching and recognition method based on high-contrast features, according to the present invention, includes: Step S1: Perform Gaussian filtering for noise reduction and histogram equalization preprocessing on the space digital X-ray image; Step S2: Extract high-contrast features using the local binary mode algorithm, obtain multi-scale features by combining different neighborhood radii and number of pixels, and then perform concatenation and fusion to form a high-contrast feature description vector; Step S3: Using the high-contrast feature description vector, image feature points are matched using the scale-invariant feature transformation algorithm, and the affine transformation model is estimated by combining the random sampling consensus algorithm to realize panoramic stitching of aerospace digital ray images. Step S4: Construct a convolutional neural network, train the model using the cross-entropy loss function, and input the stitched panoramic image into the trained model to complete intelligent target recognition.

[0008] Preferably, the standard deviation σ of the Gaussian filter in step S1 ranges from 1 to 2, the Gaussian kernel is an odd-order square kernel, and when σ is between 1.0 and 1.5, it corresponds to a 3×3 Gaussian kernel, and when σ is between 1.5 and 2.0, it corresponds to a 5×5 Gaussian kernel.

[0009] Preferably, the histogram equalization in step S1 is an adaptive histogram equalization specific to the aerospace scene, which specifically includes: image region segmentation, grayscale range limitation, block adaptive equalization, sub-block grayscale fusion, secondary contrast fine-tuning, and effective region restoration.

[0010] Preferably, the image region segmentation is based on a grayscale threshold screening. The grayscale of the background of the aerospace ray image is concentrated in the range of 0 to 50, and the grayscale of the component region is concentrated in the range of 50 to 200. Only the effective region of the component is equalized. The grayscale range limitation limits the equalized grayscale range to [20, 235], and the secondary contrast fine-tuning controls the contrast enhancement coefficient to 1.0 to 1.2.

[0011] Preferably, in step S2, the neighborhood radius R of the local binary mode algorithm is 1 and 2, the number of neighborhood pixels P is 8 and 16, and the multi-scale features include four sets of scales: R=1, P=8, R=1, P=16, R=2, P=8, and R=2, P=16. The feature histograms of each set are concatenated in this order to form a feature description vector.

[0012] Preferably, in step S3, when matching feature points, the threshold range for the ratio of nearest neighbor distance to second nearest neighbor distance is 0.7 to 0.9. The affine transformation model is used to map the coordinates of the original image to the coordinates of the transformed image, thereby enabling panoramic stitching of multiple images.

[0013] Preferably, the model training in step S4 uses the cross-entropy loss function and the backpropagation algorithm.

[0014] A digital ray image stitching and recognition system based on high-contrast features, according to the present invention, includes: Preprocessing module: used for Gaussian filtering and adaptive histogram equalization preprocessing of aerospace digital X-ray images; Feature extraction module: used to extract multi-scale high-contrast features using the local binary mode algorithm, and to concatenate and fuse the multi-scale features to output a high-contrast feature description vector; Panoramic stitching module: Used to perform feature point matching using the scale-invariant feature transformation algorithm, combined with the random sampling consensus algorithm to estimate the affine transformation model, to complete the panoramic stitching of aerospace digital ray images; Intelligent recognition module: used to build a convolutional neural network, train the model with cross-entropy loss function, and perform intelligent target recognition on the stitched panoramic image, outputting the recognition results.

[0015] Preferably, it also includes a data storage module for storing the preprocessed image, extracted high-contrast features, stitched panoramic image, trained convolutional neural network model, and target recognition results.

[0016] Preferably, the preprocessing module is further used to divide the effective area of ​​the component into blocks and perform grayscale fusion of the sub-blocks to eliminate grayscale abrupt changes between sub-blocks.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention employs Gaussian filtering for noise reduction and histogram equalization for preprocessing, combined with a multi-scale local binary mode algorithm to extract high-contrast features. Compared with the simple contrast adjustment methods of existing technologies, this invention achieves a balance between noise reduction and detail preservation, effectively improving the contrast and feature recognition of aerospace DR images, and solving the problem of low distinction between defects and background in traditional methods.

[0018] 2. Based on the extracted high-contrast features, this invention uses a scale-invariant feature transformation algorithm to match feature points and combines a random sampling consensus algorithm to estimate the affine transformation model to complete panoramic stitching. Compared with existing stitching schemes that are not adapted to the characteristics of DR images, this invention significantly reduces the feature mismatch rate, improves stitching accuracy and stability, generates panoramic images without misalignment or ghosting, and fully preserves defect details.

[0019] 3. This invention constructs a multi-convolutional layer CNN model, which is trained with the cross-entropy loss function to perform global intelligent recognition on the stitched panoramic image. Compared with the existing technology based on local image recognition, it makes full use of the global field of view and high-contrast defect features of the panoramic image, significantly improving the recognition accuracy of small internal defects of aerospace components and meeting the high-precision requirements of aerospace non-destructive testing. Attached Figure Description

[0020] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the working method of the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0022] Example 1 According to the present invention, a digital ray image stitching and recognition method based on high contrast features is provided, such as... Figure 1 As shown, it includes: Step S1: Perform Gaussian filtering for noise reduction and histogram equalization preprocessing on the spaceborne digital X-ray image. The preprocessing uses Gaussian filtering to remove image noise and histogram equalization to enhance image contrast. The Gaussian filtering formula is as follows: (1) in, G(x,y) This represents the value of the Gaussian function at coordinates (x, y), used to calculate the filtering weight at the corresponding position in the image, and to determine the degree of change of the pixel value at that position during the filtering process; x, y Represents coordinate values ​​on the image plane, used to locate pixels in the image; σ This represents the standard deviation and controls the width of the Gaussian function. x 0 、y 0 represents the coordinates of the image center. The standard deviation... σ The value range is 1 to 2.

[0023] Considering the noise characteristics of aerospace digital X-ray images and matching the preferred parameter of standard deviation σ ranging from 1 to 2 as specified in the instruction manual, an odd-order square Gaussian kernel is adopted. The specific size and adaptation relationship with σ are as follows, which is the optimal choice for aerospace scenarios: Table 1 Gaussian kernel sizes for different standard deviations

[0024] Aerospace digital X-ray images suffer from uneven background grayscale, extremely low contrast between the target and background, and a narrow dynamic range of grayscale in component areas. Conventional histogram equalization can easily lead to noise amplification and the obscuring of defect features. Therefore, an adaptive histogram equalization step specifically designed for aerospace scenarios was developed to enhance contrast and adapt to the inspection requirements of aerospace components. The specific steps are as follows: Step S101: Image Region Segmentation. The input aerospace digital ray image is segmented into aerospace component regions and background regions (based on initial screening using image grayscale thresholds; the background grayscale of aerospace ray images is usually concentrated in the low grayscale range of 0-50, while the grayscale of component regions is 50-200). Histogram equalization is performed only on the effective regions of the components to avoid amplifying background noise.

[0025] Step S102: Gray-scale range limitation. Based on the gray-scale characteristics of aerospace digital ray images, the gray-scale range for equalization is limited to [20, 235] to prevent extreme gray-scale values ​​from interfering with the equalization effect and to ensure the effective expansion of gray-scale levels within the components.

[0026] Step S103: Block Adaptive Equalization. The effective area of ​​the component is divided into 8×8 / 16×16 non-overlapping sub-blocks (adjusted according to the image resolution; the commonly used resolution for aerospace ray images is 2048×2048, with 16×16 sub-blocks being preferred). Histogram equalization is performed on each sub-block individually to solve the problem of uneven background grayscale in aerospace ray images and avoid the loss of local features caused by overall equalization.

[0027] Step S104: Sub-block grayscale fusion. The sub-blocks after block equalization are smoothly fused in grayscale. Bilinear interpolation is used to eliminate abrupt changes in grayscale between sub-blocks, ensuring the continuity of grayscale in the component area and preventing splicing marks from covering up minor defects.

[0028] Step S105: Secondary contrast fine-tuning. The fused image is subjected to grayscale stretching fine-tuning, mapping the grayscale dynamic range to [0,255], and the contrast enhancement coefficient is controlled between 1.0 and 1.2 to avoid excessive enhancement that would amplify noise, while ensuring that the grayscale difference between the defects and the substrate is significant.

[0029] Step S106: Effective Area Restoration. The equalized component area is stitched together with the original background area to restore the complete aerospace digital ray preprocessed image, ensuring image integrity.

[0030] The contrast enhancement coefficient after image preprocessing is controlled within the range of 1.0 to 1.2, and adaptive adjustments are made based on different objects and specific parts.

[0031] Step S2: Extract high-contrast features using the Local Binary Mode (LBMT) algorithm. Step S2 includes obtaining multi-scale features through different combinations of neighborhood radii and pixel counts. The LBMT formula is as follows: (2) in, Indicates a neighborhood radius of R The number of neighboring pixels is P In the case of a certain value, the local binary pattern encoded value of the center pixel is used to describe the grayscale relationship between the center pixel and its neighboring pixels. P Indicates the number of neighboring pixels; R Indicates the neighborhood radius; Represents the grayscale value of neighboring pixels. p From 0 to P 1 , representing the grayscale value of pixels at different locations within the neighborhood. Indicates the grayscale value of the center pixel; s(x) For a sign function, when x When ≥0, s(x) =1; when x When <0, s(x) =0. By combining different neighborhood radii and pixel numbers, high-contrast features at multiple scales can be obtained; neighborhood radius R The values ​​include 1 and 2, representing the number of neighboring pixels. P The possible values ​​are 8 and 16.

[0032] The calculation order of the local binary model formula is as follows: 1) Calculate the small-scale basic feature (R1P8). Using a neighborhood radius R=1 and a neighborhood number of pixels P=8 as parameters, calculate the LBP encoding value pixel by pixel for the preprocessed aerospace digital ray image. This scale is the smallest basic scale, which can accurately capture subtle local features such as microcracks and micropores on the surface of aerospace components. It is the basis for subsequent multi-scale features and the core feature scale for defect identification in aerospace nondestructive testing.

[0033] 2) Calculate the same-radius extended pixel features (R1P16). Keeping the neighborhood radius R=1 unchanged, the number of neighborhood pixels is expanded to P=16, and the LBP encoding value is calculated pixel by pixel. This scale, based on the subtle features of R1P8, supplements the surrounding gray-level correlation features under the same radius, weakens the local noise interference of aerospace images, strengthens the gray-level contour features of small defects, and avoids the feature randomness of single-pixel scale.

[0034] 3) Expand the neighborhood radius to R=2, restore the number of neighborhood pixels to P=8, and calculate the LBP encoding value pixel by pixel. This scale is a large-scale basic scale, which can capture mesoscopic structural features such as welds, bolt holes, and pipelines of aerospace components, making up for the problem of insufficient capture of overall structural features by small-radius scales, and providing structural support for feature point matching in subsequent image stitching.

[0035] 4) Calculate the large-radius extended pixel feature (R2P16). Keeping the neighborhood radius R=2 unchanged, the number of neighborhood pixels is expanded to P=16, and the LBP encoding value is calculated pixel by pixel. This scale is the maximum fusion scale, which can capture the macroscopic overall grayscale distribution features of aerospace components, strengthen the grayscale boundary features of different structural regions, improve the robustness of features, and adapt to the low contrast characteristics of aerospace digital ray images.

[0036] 5) Multi-scale LBP feature concatenation and fusion. Following the calculation order described above, the LBP encoded values ​​at the four scales are statistically analyzed into single-scale feature histograms. Then, the feature histograms are concatenated in one dimension in the order of R1P8→R1P16→R2P8→R2P16 to form the final multi-scale LBP high-contrast feature description vector, which is used for feature point extraction and matching in subsequent panoramic image stitching.

[0037] Step S3: Utilizing the high-contrast features, image feature points are matched using a scale-invariant feature transform algorithm, and the affine transformation model is estimated using a random sample consensus algorithm to achieve panoramic image stitching. Specifically, using the extracted high-contrast features, a scale-invariant feature transform algorithm is employed to extract feature points and calculate descriptors in the images to be stitched. Initial matching is performed by comparing the Euclidean distances of feature point descriptors. The ratio of the nearest neighbor distance to the second nearest neighbor distance has a threshold range of 0.7 to 0.9, which filters out reliable matching point pairs. The Euclidean distance formula is as follows: (3) in, d The distance represents the Euclidean distance between two feature point descriptors, which is used to measure the similarity between feature points. The smaller the distance, the more similar the feature points are. n The dimension of a feature point descriptor is the number of feature dimensions contained in each feature point descriptor. Different feature extraction algorithms... n The values ​​can be different. This represents the i-th dimension component of the first feature point descriptor. This represents the i-th dimension component of the second feature point descriptor.

[0038] Based on the matching point pairs, the affine transformation model between images is estimated using the random sampling consensus algorithm to achieve panoramic stitching. The formula is as follows: (4) in, x, y Represents the original image coordinates, used to indicate the position of a pixel in the original image; x′, y′ Represents the coordinates of the transformed image; after the affine transformation, the position of the pixel in the original image in the new image. , , , Represents the transformation matrix parameters; , These represent translation parameters, indicating the amount of translation in the x-axis and y-axis directions, respectively.

[0039] Step S4: Construct a convolutional neural network (CNN) and train the model using the cross-entropy loss function to perform intelligent object recognition on the stitched panoramic image. For the stitched panoramic image, a CNN model containing convolutional layers, pooling layers, and fully connected layers is constructed. The model consists of 5 convolutional layers, 3 pooling layers, and 2 fully connected layers. The kernel sizes of the convolutional layers are 3×3 and 5×5, respectively, and the pooling layers use 2×2 max pooling. The number of neurons in the fully connected layers is 512, and the number of classes is [not specified in the original text].

[0040] The formula for feature extraction using convolution operations is as follows: (5) in, feature i,j This represents the value of the feature map at (i,j) after convolution; i, j Represents the coordinates of the feature map, used to locate the pixel position in the feature map after convolution; M, N This indicates the size of the convolution kernel, that is, the dimensions of the convolution kernel in the horizontal and vertical directions; input i+m,j+n This represents the value of the input image at (i+m,j+n); kernel m,n The value of the convolution kernel at (m,n) represents the value of the kernel. The convolution kernel is a fixed-size matrix that extracts features from the image by performing a convolution operation with the input image. b represents the bias term.

[0041] Using a large number of spaceborne digital ray images, a model was trained using the cross-entropy loss function and backpropagation algorithm. Panoramic images were then input into the trained model for target recognition, yielding the recognition results. The formula for the cross-entropy loss function is as follows: (6) in, L This represents the cross-entropy loss value, which measures the degree of difference between the model's prediction and the true label. The smaller the loss value, the closer the model's prediction is to the true label. C Indicates the number of categories; y i Indicates the first in the real label i The value of the class can be either 0 or 1, when the sample belongs to the first class. i When class, y i =1; otherwise y i =0; p i The model predicts the first... i The probability of a class ranges from 0 to 1. When training the model using the cross-entropy loss function and backpropagation algorithm, the learning rate ranges from 0.0005 to 0.002, and the number of iterations ranges from 80 to 120.

[0042] This invention continuously adjusts the model parameters through a backpropagation algorithm to minimize the loss function, thereby improving the model's recognition accuracy. Further, the specific parameters of this invention's digital ray image stitching and recognition method based on high-contrast features are described below: First, image preprocessing. The acquired aerospace digital X-ray images are first subjected to Gaussian filtering for noise reduction. The standard deviation of the Gaussian kernel is set to σ=1.5 to effectively remove noise interference while preserving as much detail as possible. Then, histogram equalization is performed to expand the grayscale range of the image to the entire grayscale interval [0,255], enhancing image contrast and making the target and background in the image clearer and more distinguishable.

[0043] Next, high-contrast feature extraction is performed. The Local Binary Pattern (LBP) algorithm is used to extract high-contrast features from the image. Neighborhood radii R=1 and R=2, and the number of neighboring pixels P=8 and P=16, respectively, are used to calculate LBP features at different scales. For each pixel, a corresponding binary code is generated according to the LBP calculation formula, and its frequency of occurrence in the image is counted to form a feature histogram. The feature histograms at different scales are concatenated to obtain a description of the high-contrast features of the image.

[0044] Next, panoramic image stitching is performed. The Scale Invariant Feature Transform (SIFT) algorithm is used to extract feature points from the two images to be stitched, and descriptors for these feature points are calculated. Initial matching is performed by comparing the Euclidean distance between the feature point descriptors, with a ratio test threshold of 0.8 set to filter out reliable matching point pairs.

[0045] Next, the Random Sample Consensus (RANSAC) algorithm is used to estimate the affine transformation model between images. Four sets of matching point pairs are randomly selected, the affine transformation matrix is ​​calculated, and then all matching points are verified, counting the number of interior points that satisfy the transformation model. After multiple iterations, the affine transformation model with the most interior points is selected as the final transformation model. Based on this model, one image is transformed to achieve stitching with another image. The above process is repeated to stitch multiple images into a panoramic image.

[0046] Finally, the intelligent recognition system constructs a convolutional neural network (CNN) model containing 5 convolutional layers, 3 pooling layers, and 2 fully connected layers. The convolutional kernel sizes are 3×3 and 5×5, respectively, and the pooling layers use 2×2 max pooling. The fully connected layers have 512 neurons and [the number of classes is missing from the original text].

[0047] 5000 labeled aerospace digital ray images were collected as a training set, and 1000 as a test set. Data augmentation, such as random rotation and flipping, was performed on the training set images to increase data diversity. A CNN model was trained using the cross-entropy loss function and stochastic gradient descent algorithm, with a learning rate of 0.001 and 100 iterations. During training, the model's accuracy on the test set was monitored in real time. Training was stopped when the accuracy no longer improved, resulting in a trained recognition model. The stitched aerospace digital ray panoramic image was input into the trained model, and the target recognition result was output. Recognition tests were conducted on digital ray images of 11 aerospace products of a certain model. The recall rate of images with defects exceeding standards was 100%, and the image recognition accuracy was consistently above 88.3%, meeting the intelligent application requirements for digital ray detection images of a certain aerospace product model. The model has passed process evaluation and is used in daily production. The detection results on different models of a certain aerospace product showed good stability and reliability, with a single image detection efficiency 150% higher than traditional methods.

[0048] This invention organically integrates multiple technologies, fully considering the compatibility, synergy, and contribution to the final goal of each technology. In the process of technology integration, this invention designs and optimizes technologies such as image preprocessing, feature extraction and stitching, and deep learning recognition, so that each technology can cooperate and work together within the same framework to achieve full-process optimization. This not only solves the problem of poor processing effect caused by the independent operation of each link in the existing technology, but also creates a brand-new and efficient digital ray panoramic image processing mode.

[0049] Example 2 The present invention also provides a digital ray image stitching and recognition system based on high contrast features. The digital ray image stitching and recognition system based on high contrast features can be implemented by executing the process steps of the digital ray image stitching and recognition method based on high contrast features. That is, those skilled in the art can understand the digital ray image stitching and recognition method based on high contrast features as a preferred embodiment of the digital ray image stitching and recognition system based on high contrast features.

[0050] According to the present invention, a digital ray image stitching and recognition system based on high-contrast features includes: a preprocessing module for performing Gaussian filtering denoising and adaptive histogram equalization preprocessing on aerospace digital ray images; the preprocessing module is also used for segmenting the effective regions of components and performing gray-level fusion of sub-blocks to eliminate gray-level abrupt changes between sub-blocks; a feature extraction module for extracting multi-scale high-contrast features using a local binary mode algorithm and performing concatenated fusion of multi-scale features to output a high-contrast feature description vector; a panoramic stitching module for matching feature points using a scale-invariant feature transformation algorithm and estimating an affine transformation model using a random sampling consensus algorithm to complete panoramic stitching of aerospace digital ray images; and an intelligent recognition module for constructing a convolutional neural network, training the model with a cross-entropy loss function, and performing intelligent target recognition on the stitched panoramic image to output the recognition result. It also includes a data storage module for storing the preprocessed image, extracted high-contrast features, the stitched panoramic image, the trained convolutional neural network model, and the target recognition result.

[0051] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0052] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A digital ray image stitching and recognition method based on high-contrast features, characterized in that, include: Step S1: Perform Gaussian filtering for noise reduction and histogram equalization preprocessing on the space digital X-ray image; Step S2: Extract high-contrast features using the local binary mode algorithm, obtain multi-scale features by combining different neighborhood radii and number of pixels, and then perform concatenation and fusion to form a high-contrast feature description vector; Step S3: Using the high-contrast feature description vector, image feature points are matched using the scale-invariant feature transformation algorithm, and the affine transformation model is estimated by combining the random sampling consensus algorithm to realize panoramic stitching of aerospace digital ray images. Step S4: Construct a convolutional neural network, train the model using the cross-entropy loss function, and input the stitched panoramic image into the trained model to complete intelligent target recognition.

2. The digital ray image stitching and recognition method based on high-contrast features according to claim 1, characterized in that, The standard deviation σ of the Gaussian filter in step S1 ranges from 1 to 2. The Gaussian kernel is an odd-order square kernel, and when σ is between 1.0 and 1.5, it corresponds to a 3×3 Gaussian kernel, and when σ is between 1.5 and 2.0, it corresponds to a 5×5 Gaussian kernel.

3. The digital ray image stitching and recognition method based on high-contrast features according to claim 1, characterized in that, The histogram equalization in step S1 is an adaptive histogram equalization specific to the aerospace scene, which includes: image region segmentation, grayscale range limitation, block adaptive equalization, sub-block grayscale fusion, secondary contrast fine-tuning, and effective region restoration.

4. The digital ray image stitching and recognition method based on high contrast features according to claim 3, characterized in that, The image region segmentation is based on the initial screening of grayscale thresholds. The grayscale of the background of the aerospace ray image is concentrated in the range of 0 to 50, and the grayscale of the component region is concentrated in the range of 50 to 200. Only the effective region of the component is equalized. The grayscale range limitation limits the equalized grayscale range to [20, 235], and the secondary contrast fine-tuning controls the contrast enhancement coefficient to 1.0 to 1.

2.

5. The digital ray image stitching and recognition method based on high contrast features according to claim 1, characterized in that, In step S2, the neighborhood radius R of the local binary mode algorithm is 1 and 2, the number of neighborhood pixels P is 8 and 16, and the multi-scale features include four sets of scales: R=1, P=8, R=1, P=16, R=2, P=8, and R=2, P=16. The feature histograms of each set are concatenated to form a feature description vector.

6. The digital ray image stitching and recognition method based on high contrast features according to claim 1, characterized in that, In step S3, when matching feature points, the threshold range for the ratio of nearest neighbor distance to second nearest neighbor distance is 0.7 to 0.

9. The affine transformation model is used to map the coordinates of the original image to the coordinates of the transformed image, thereby enabling panoramic stitching of multiple images.

7. The digital ray image stitching and recognition method based on high contrast features according to claim 1, characterized in that, In step S4, the model training uses the cross-entropy loss function and the backpropagation algorithm.

8. A digital ray image stitching and recognition system based on high contrast features, characterized in that, include: Preprocessing module: used for Gaussian filtering and adaptive histogram equalization preprocessing of aerospace digital X-ray images; Feature extraction module: used to extract multi-scale high-contrast features using the local binary mode algorithm, and to concatenate and fuse the multi-scale features to output a high-contrast feature description vector; Panoramic stitching module: Used to match feature points using the scale-invariant feature transform algorithm and estimate the affine transformation model using the random sampling consensus algorithm to complete the panoramic stitching of aerospace digital ray images; Intelligent recognition module: used to build a convolutional neural network, train the model with cross-entropy loss function, and perform intelligent target recognition on the stitched panoramic image, outputting the recognition results.

9. The digital ray image stitching and recognition system based on high contrast features according to claim 8, characterized in that, It also includes a data storage module for storing preprocessed images, extracted high-contrast features, stitched panoramic images, trained convolutional neural network models, and target recognition results.

10. The digital ray image stitching and recognition system based on high contrast features according to claim 8, characterized in that, The preprocessing module is also used to divide the effective area of ​​the component into blocks and perform grayscale fusion of sub-blocks to eliminate grayscale abrupt changes between sub-blocks.

Images