A Deep Learning-Based Visual Inspection and Grading System for Concrete Texture Quality

Abstract

This invention provides a deep learning-based visual inspection and grading system for concrete roughening quality, applicable to the field of concrete roughening technology. The system includes a data acquisition module, a dynamic preprocessing module, a deep learning model recognition module, and a decision interaction module. It converts three-dimensional spatial coordinate information into concrete roughening quality evaluation indicators for quantitative quality grading, achieving accurate classification of under-roughening, acceptable, and over-roughening. It employs Brenner function verification and liquid focusing lens control to achieve automatic focusing on rough surfaces, improving image clarity and accelerating shooting speed. The model, pre-trained on ImageNet and optimized with Focal Loss, can quickly output quality levels after a single shot. Combined with improved SIFT feature matching, it enables image stitching and rework area annotation.

Description

Technical Field

[0001] This invention relates to the field of concrete roughening technology, specifically to a deep learning-based visual inspection and grading system for concrete roughening quality. Background Technology

[0002] In concrete pouring, it is usually impossible to complete the entire pour in one go; it must be done in stages. Inevitably, the load-bearing capacity and durability of the interface between new and old concrete must be considered. For concrete structures in special locations, the impermeability and erosion resistance of this interface must also be taken into account. Therefore, before pouring new concrete at the interface, the old concrete needs to be roughened or roughened to increase its roughness and allow for better bonding between the new and old concrete. Traditional roughening methods require workers to use roughening tools, which is time-consuming and labor-intensive, and has gradually been replaced by high-pressure water jet roughening. However, roughening requires the old concrete to have solidified to a certain strength but before it is completely hardened, which places relatively high demands on processing time.

[0003] To ensure reliable bonding between new and old concrete, current standards stipulate that, quantitatively, the depth of the roughened surface should be no less than 6mm, and qualitatively, the roughened surface should slightly expose coarse sand. Generally, the roughness of the concrete surface is determined by construction personnel through direct observation based on their experience. For projects requiring precise quantitative analysis, methods such as the sand filling method, roughness meter method, fractal dimension method, needle contact method, and small iron bead measurement method have been used. However, these methods are all complex to operate and can significantly impact construction efficiency in practical projects. Additionally, there are methods based on 3D point cloud scanning for roughness assessment of the concrete surface after roughening; however, the scanning and data processing efficiency of this method is low. Therefore, there is an urgent need for a rapid and accurate method for detecting the quality of concrete roughening. Summary of the Invention

[0004] The purpose of this invention is to solve the problems of low efficiency in existing concrete roughening quality inspection and its impact on on-site construction progress. It proposes a deep learning-based visual inspection and grading system for concrete roughening quality, which can be widely applied in the field of concrete roughening technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A deep learning-based visual inspection and grading system for concrete surface roughness quality includes:

[0007] The data acquisition module uses an autofocus industrial camera paired with a ring light source to acquire raw image data of the rough surface of the concrete construction joint after roughening.

[0008] The dynamic preprocessing module is used to optimize the quality of raw image data in real time;

[0009] The deep learning model recognition module identifies the quality level of the rough surface of the concrete construction joint after the concrete has been roughened by a visual inspection model for concrete roughening quality.

[0010] The decision interaction module is used to display real-time detection results, stitch together the original image data to form an overall image of the scanned area, and mark the rework areas.

[0011] As a preferred technical solution of the present invention, the autofocus industrial camera achieves autofocus based on a deep learning-based defocus distance prediction algorithm. The specific implementation process includes model architecture design, dataset construction, model optimization training, and focus strategy implementation.

[0012] The model architecture is designed to extract image features using the lightweight CNN network ShuffleNetV2 and add 3 layers of MLP to map the features output by ShuffleNetV2 to the defocus distance.

[0013] The dataset was constructed using a liquid focusing lens. The refractive index of the liquid was changed by adjusting the voltage value with a fixed step size, and image data at different defocus distances were collected. The image dataset collected at each data acquisition point covered the range from severe defocus to optimal focus. The Brenner sharpness evaluation function was used to determine the optimal focus position, and the optimal focus position was marked with a defocus distance of 0. The defocus distance l of other images was calculated using formula (1).

[0014] l = |V i -V b | (1)

[0015] In the formula, V i - The voltage corresponding to the liquid focusing lens during the i-th shot, V b - The voltage corresponding to the liquid focusing lens during optimal focusing and shooting; Divide all image datasets collected from different data acquisition points into training set, validation set and test set in a 7:2:1 ratio;

[0016] The model was optimized and trained using the Huber loss function, L2 regularization, and Adam optimizer to train the defocus distance prediction model with an initial learning rate of 0.0001. The mean absolute error (MAE) and coefficient of determination (R²) were used. 2 Measure the predictive performance of the model;

[0017] The focusing strategy is implemented based on the predicted defocus distance l. p Divided into three levels, l p ∈(10,+∞) represents a large defocus area; a large step size is used to quickly approach the optimal focus position. p∈(1,10] represents the mid-focus area, using a medium step size to balance speed and accuracy, l p ∈(0,1] represents small defocus, which is finely adjusted using micro-steps; with one image acquisition, the model predicts the defocus distance l. p The system then dynamically matches the step size and moves directly to the target area. It then uses the Brenner function to verify the focus quality. If the quality is not up to standard, it triggers a second small defocus adjustment.

[0018] As a preferred technical solution of the present invention, the real-time optimization includes using a multi-scale Retinex algorithm to decompose the incident light and reflected light components, and using multi-scale spatial convolution of Gaussian filter kernels to eliminate illumination unevenness.

[0019] As a preferred technical solution of the present invention, the training process of the visual inspection model for concrete roughening quality includes collecting training sample datasets, data preprocessing, model architecture design, and model training.

[0020] A training sample dataset was collected, and the shooting distance between the shooting device and the rough surface of the concrete construction joint was fixed so that the shooting range was consistent with the single scan range of the 3D line laser scanner. Each scan corresponds to taking an image of the rough surface of the concrete construction joint. The 3D spatial coordinate information of the rough surface of the concrete construction joint after the concrete structure has been roughened was collected by the 3D line laser scanner. The quality level was evaluated based on the 3D spatial coordinate information to obtain the quality level of the corresponding image.

[0021] Data preprocessing employs a multi-scale Retinex algorithm to decompose incident and reflected light components, eliminates illumination unevenness through multi-scale spatial convolution with Gaussian filter kernels, and enhances data diversity by applying rotation, flipping, and brightness / contrast jitter. A hybrid oversampling and undersampling strategy is used to generate synthetic samples based on feature space interpolation, and the K-nearest neighbor rule is applied to filter boundary samples for interpolation. A dynamic weight allocation mechanism is introduced to adjust the regional weights of generated samples according to class density, prioritizing the addition of synthetic samples in classification boundary regions. Generative adversarial networks are combined to augment data in the latent space, and the discriminator-generator game improves sample authenticity. Noise filtering is performed simultaneously, using the edit nearest neighbor algorithm to remove abnormal samples in overlapping regions.

[0022] The model architecture design uses an improved ResNet-34 as the backbone network for the feature extraction layer, embedding an ECA attention module to enhance feature extraction; a multi-scale fusion layer is constructed, introducing bidirectional cross-layer connections and deformable convolution modules; a hierarchical decision layer integrates a spatial pyramid pooling layer to distinguish between three categories of concrete roughening quality: under-brushing, qualified, and over-brushing.

[0023] For model training, the backbone network was initialized based on ImageNet pre-trained weights, the parameters of the first three convolutional layers were frozen, and the learning rate of the high-level modules was fine-tuned using cosine annealing. Then, the parameters of the first three convolutional layers were unfrozen, Focal Loss was introduced to balance the distribution of undershot / overshot samples, and γ=2.0 was set to suppress the weights of easily classified samples. The AdamW optimizer was used for training, and the learning rate adopted a phased decay strategy. Early stopping detection on the validation set was performed every 10 epochs, and model rollback was triggered when the loss did not decrease after three consecutive validations.

[0024] As a preferred embodiment of the present invention, the quality level evaluation is as follows:

[0025] Q = w1H1 + w2S1 + w3N1 (2)

[0026] In the formula, Q represents the concrete roughening quality evaluation index, w1, w2, and w3 represent weighting coefficients, H1 represents the normalized maximum vertical height difference between adjacent peaks and troughs, S1 represents the normalized maximum peak and trough slope, and N1 represents the normalized number of peaks per unit area; the normalization is based on the maximum and minimum reference values ​​of each index.

[0027]

[0028]

[0029] In the formula, H represents the maximum vertical height difference between adjacent peaks and troughs in a single scan area (mm), S represents the maximum slope of the peak and trough in a single scan area, and N represents the number of peaks per unit area (z). p,i - Height of the i-th peak, mm, z v,i - The height of the trough adjacent to the i-th peak, mm, (x mp y mp , z mp - Coordinates of the maximum peak, (x mv y mv , z mv ) - Coordinates of the maximum trough, n - Number of peaks in a single scan area, s - Area of ​​a single scan area; Maximum and minimum reference values ​​are statistically analyzed based on historical quality test results; When Q≤0.2, the quality level is evaluated as undershoot; when 0.2<Q≤0.8, the quality level is evaluated as qualified; when Q>0.8, the quality level is evaluated as overshoot.

[0030] As a preferred technical solution of the present invention, the stitching of the original image data adopts an improved SIFT feature matching combined with the RANSAC algorithm, sets a matching point threshold of ≥50 pairs, and the homography matrix reprojection error is ≤1.5 pixels.

[0031] The beneficial effects of this invention are: 1. Quantitative quality grading: By converting three-dimensional spatial coordinate information into concrete roughening quality evaluation indicators, accurate classification of under-scoring / qualified / over-scoring is achieved; 2. The use of Brenner function verification and liquid focusing lens control enables automatic focusing on rough surfaces, improving image clarity and accelerating shooting speed; 3. The model, pre-trained on ImageNet and optimized with Focal Loss, can quickly output quality levels after a single shot, and combined with improved SIFT feature matching, achieves image stitching and rework area annotation. Attached Figure Description

[0032] Figure 1 This is a flowchart of the visual inspection and grading system for concrete roughening quality based on deep learning, which is the workflow of the present invention. Detailed Implementation

[0033] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments given herein are for illustration and explanation only and are not intended to limit the invention. It should be noted that many specific details are set forth in the following description to provide a thorough understanding of the invention; however, the invention may have other embodiments and modifications thereof. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0034] Example 1: A deep learning-based visual inspection and grading system for concrete surface roughness, comprising a data acquisition module, a dynamic preprocessing module, a deep learning model recognition module, and a decision interaction module.

[0035] Module 1, the data acquisition module, uses an autofocus industrial camera paired with a ring light source to acquire raw image data of the roughened surface of concrete construction joints after roughening. The autofocus industrial camera achieves automatic focusing based on a deep learning-based defocus distance prediction algorithm, enabling rapid and accurate focusing on rough surfaces, ensuring image clarity and image capture speed. The specific implementation process includes model architecture design, dataset construction, model optimization and training, and focusing strategy implementation.

[0036] The model architecture design utilizes the lightweight CNN network ShuffleNetV2 to extract image features and adds 3 layers of MLP to map the features output by ShuffleNetV2 to the defocus distance.

[0037] Dataset construction: A liquid focusing lens is used. The voltage value is adjusted by a fixed step of 0.2V to change the refractive index of the liquid, and image data at different defocus distances are collected. The image dataset collected at each data acquisition point covers the range from severe defocus to optimal focus. The Brenner sharpness evaluation function is used to determine the optimal focus position, and the defocus distance at the optimal focus position is marked as 0. The defocus distance l of other images is calculated by formula (1).

[0038] l = |V i - V b | (1)

[0039] In the formula, V i - The voltage corresponding to the liquid focusing lens during the i-th shooting, V b - The voltage corresponding to the liquid focusing lens during the optimal focus shooting; The dataset is divided by acquisition points to ensure that the training set / validation set / test set = 7:2:1, avoiding misallocation of data at the same point to different datasets.

[0040] Model optimization training: The defocus distance prediction model is trained using the Huber loss function, L2 regularization term, and Adam optimizer. The initial learning rate is 0.0001, and it decays by 50% every 20 epochs. Training is terminated when the validation loss does not decrease for 5 consecutive times; The mean absolute error MAE and the coefficient of determination R 2 are used to measure the prediction performance of the model.

[0041] Focusing strategy implementation: The current frame image is collected and input into the defocus distance prediction model to obtain l p , and the voltage adjustment step ΔV is selected according to the l p level:

[0042] If l p ∈(10, +∞) is large defocus, a large step ΔV = 1V is used to quickly approach the optimal focus position;

[0043] If l p ∈(1, 10] is medium defocus, a medium step ΔV = 0.5V is used to balance speed and accuracy;

[0044] If l p ∈(0, 1] is small defocus, a micro step ΔV = 0.1V is used for fine adjustment.

[0045] After adjustment, a new image is collected, and the Brenner value is calculated to verify the sharpness. If Brenner ≥ threshold T, the focusing is completed. If Brenner < T: Trigger secondary small defocus fine-tuning until the threshold T is reached or the maximum adjustment次数 is exceeded; Record the final V b value, further update the dataset, and train the defocus distance prediction model.

[0046] Module 2, the dynamic preprocessing module, is used to optimize the quality of the original image data in real time. Real-time optimization includes using the multi-scale Retinex algorithm to decompose the incident and reflected light components, and eliminating illumination unevenness through multi-scale spatial convolution of Gaussian filter kernels.

[0047] Module 3, the deep learning model recognition module, identifies the quality level of the rough surface of concrete construction joints after roughening using a visual inspection model for concrete roughening quality. The training process of the concrete roughening quality visual inspection model includes collecting training sample datasets, data preprocessing, model architecture design, and model training.

[0048] A training sample dataset was collected, and the shooting distance between the shooting device and the rough surface of the concrete construction joint was fixed so that the shooting range was consistent with the single scan range of the 3D line laser scanner. Each scan corresponds to taking an image of the rough surface of the concrete construction joint. The 3D spatial coordinate information of the rough surface of the concrete construction joint after the concrete structure has been roughened was collected by the 3D line laser scanner. The quality level of the corresponding image was obtained by evaluating the quality level based on the 3D spatial coordinate information.

[0049] Data preprocessing employs a multi-scale Retinex algorithm to decompose incident and reflected light components, eliminates illumination unevenness through multi-scale spatial convolution with Gaussian filter kernels, and enhances data diversity by applying rotation, flipping, and brightness / contrast jitter. A hybrid oversampling and undersampling strategy is used to generate synthetic samples based on feature space interpolation, and the K-nearest neighbor rule is applied to filter boundary samples for interpolation. A dynamic weight allocation mechanism is introduced to adjust the regional weights of generated samples according to class density, prioritizing the addition of synthetic samples in classification boundary regions. Generative adversarial networks are combined to augment data in the latent space, and the discriminator-generator game improves sample authenticity. Noise filtering is performed simultaneously, using the edit nearest neighbor algorithm to remove abnormal samples in overlapping regions.

[0050] In the model architecture design, the feature extraction layer uses an improved ResNet-34 as the backbone network, embedding an ECA attention module to enhance feature extraction, and the first layer convolutional kernel is adjusted to a 5×5 size; a multi-scale fusion layer is constructed, introducing bidirectional cross-layer connections and deformable convolutional modules; the hierarchical decision layer integrates a spatial pyramid pooling layer to distinguish between three categories of concrete roughening quality: under-brushing, qualified, and over-brushing.

[0051] For model training, the backbone network was initialized based on ImageNet pre-trained weights, the parameters of the first three convolutional layers were frozen, and the learning rate of the high-level modules was fine-tuned using cosine annealing. Then, the parameters of the first three convolutional layers were unfrozen, Focal Loss was introduced to balance the distribution of undershot / overshot samples, and γ=2.0 was set to suppress the weights of easily classified samples. The AdamW optimizer was used for training, and the learning rate adopted a phased decay strategy. Early stopping detection on the validation set was performed every 10 epochs, and model rollback was triggered when the loss did not decrease after three consecutive validations.

[0052] When collecting training sample datasets, it is necessary to evaluate the quality level based on three-dimensional spatial coordinate information. The specific evaluation formula is as follows:

[0053] Q = w1H1 + w2S1 + w3N1 (2)

[0054] In the formula, Q represents the concrete roughening quality evaluation index, w1, w2, and w3 represent weighting coefficients, H1 represents the normalized maximum vertical height difference between adjacent peaks and troughs, S1 represents the normalized maximum peak and trough slope, and N1 represents the normalized number of peaks per unit area; the normalization is based on the maximum and minimum reference values ​​of each index.

[0055]

[0056]

[0057] In the formula, H represents the maximum vertical height difference between adjacent peaks and troughs in a single scan area (mm), S represents the maximum slope of the peak and trough in a single scan area, and N represents the number of peaks per unit area (z). p,i - Height of the i-th peak, mm, z v,i - The height of the trough adjacent to the i-th peak, mm, (x mp y mp , z mp - Coordinates of the maximum peak, (x mv y mv , z mv ) - Coordinates of the maximum trough, n - Number of peaks in a single scan area, s - Area of ​​a single scan area; Maximum and minimum reference values ​​are statistically analyzed based on historical quality test results; When Q≤0.2, the quality level is evaluated as undershoot; when 0.2<Q≤0.8, the quality level is evaluated as qualified; when Q>0.8, the quality level is evaluated as overshoot.

[0058] Module four, the decision interaction module, is used to display real-time detection results. After stitching the original image data to form a complete image of the scanned area, it marks the areas requiring rework. The stitching of the original image data uses an improved SIFT feature matching combined with the RAN SAC algorithm, setting a matching point threshold of ≥50 pairs and a homography matrix reprojection error of ≤1.5 pixels.

[0059] In summary, the deep learning-based visual inspection and grading system for concrete surface roughening quality of this invention has the characteristics of high detection efficiency and accurate evaluation in the field of concrete surface roughening technology.

[0060] It should be understood that the above embodiments are one or more embodiments of the present invention, and there are many other embodiments and variations based on the present invention; any variations and modifications made by those skilled in the art through the present invention without making pioneering innovations are all within the protection scope of the present invention.

Claims

1. A deep learning-based visual inspection and grading system for the surface roughness of concrete, characterized in that, include: The data acquisition module uses an autofocus industrial camera paired with a ring light source to acquire raw image data of the rough surface of the concrete construction joint after roughening. The dynamic preprocessing module is used to optimize the quality of raw image data in real time; The deep learning model recognition module identifies the quality level of the rough surface of the concrete construction joint after the concrete has been roughened by a visual inspection model for concrete roughening quality. The decision interaction module is used to display real-time detection results, stitch together the original image data to form an overall image of the scanned area, and mark the rework areas. The training process of the concrete roughening quality visual inspection model includes collecting training sample datasets, data preprocessing, model architecture design, and model training. A training sample dataset was collected, and the shooting distance between the shooting device and the rough surface of the concrete construction joint was fixed so that the shooting range was consistent with the single scan range of the 3D line laser scanner. Each scan corresponds to taking an image of the rough surface of the concrete construction joint. The 3D spatial coordinate information of the rough surface of the concrete construction joint after the concrete structure has been roughened was collected by the 3D line laser scanner. The quality level was evaluated based on the 3D spatial coordinate information to obtain the quality level of the corresponding image. Data preprocessing employs a multi-scale Retinex algorithm to decompose incident and reflected light components, eliminates illumination unevenness through multi-scale spatial convolution with Gaussian filter kernels, and enhances data diversity by applying rotation, flipping, and brightness / contrast jitter. A hybrid oversampling and undersampling strategy is used to generate synthetic samples based on feature space interpolation, and the K-nearest neighbor rule is applied to filter boundary samples for interpolation. A dynamic weight allocation mechanism is introduced to adjust the regional weights of generated samples according to class density, prioritizing the addition of synthetic samples in classification boundary regions. Generative adversarial networks are combined to augment data in the latent space, and the discriminator-generator game improves sample authenticity. Noise filtering is performed simultaneously, using the edit nearest neighbor algorithm to remove abnormal samples in overlapping regions. The model architecture design uses an improved ResNet-34 as the backbone network for the feature extraction layer, embedding an ECA attention module to enhance feature extraction; a multi-scale fusion layer is constructed, introducing bidirectional cross-layer connections and deformable convolution modules; a hierarchical decision layer integrates a spatial pyramid pooling layer to distinguish between three categories of concrete roughening quality: under-brushing, qualified, and over-brushing. Model training involves initializing the backbone network using ImageNet pre-trained weights, freezing the parameters of the first three convolutional layers, and fine-tuning the high-level modules using cosine annealing learning rate. Then, the parameters of the first three convolutional layers are unfrozen, and Focal Loss is introduced to balance the distribution of undershot / overshot samples. γ =2.0 Suppress the weights of easily classified samples; use the AdamW optimizer for training, and adopt a phased decay strategy for the learning rate. Perform early stopping detection on the validation set every 10 epochs. Trigger model rollback when the loss does not decrease after 3 consecutive validation cycles. The quality level evaluation is as follows: （1） In the formula, Q -Evaluation indicators for the quality of concrete roughening. w 1. w 2. w 3-Weighting coefficient, H 1- Normalized maximum vertical height difference between adjacent peaks and troughs S 1- Normalized maximum peak and maximum trough slope, N 1 - Normalized number of peaks per unit area; The normalization is performed based on the maximum and minimum reference values ​​of each indicator; The maximum and minimum reference values ​​are statistically based on historical quality test results; when Q When ≤0.2, the quality level is evaluated as undershoot; when 0.2 < Q When ≤0.8, the quality level is evaluated as qualified; when Q When the value is greater than 0.8, the quality level is evaluated as overshoot.

2. The deep learning-based visual inspection and grading system for concrete surface roughening quality according to claim 1, characterized in that: The autofocus industrial camera achieves autofocus based on a deep learning-based defocus distance prediction algorithm. The specific implementation process includes model architecture design, dataset construction, model optimization and training, and autofocus strategy implementation. The model architecture is designed to extract image features using the lightweight CNN network ShuffleNetV2 and add 3 layers of MLP to map the features output by ShuffleNetV2 to the defocus distance. Dataset construction employed a liquid focusing lens, adjusting the voltage value in fixed steps to change the liquid's refractive index, and acquiring image data at different defocus distances. The image dataset collected at each data acquisition point covered the range from severe defocus to optimal focus. The Brenner sharpness evaluation function was used to determine the optimal focus position, marked with a defocus distance of 0. The defocus distances of other images were then... l Calculated using formula (1), （2） In the formula, V i -No. i The voltage corresponding to the liquid focusing lens during the next shot. V b - The voltage corresponding to the liquid focusing lens during optimal focusing and shooting; All image datasets collected from different data collection points were divided into training set, validation set and test set in a 7:2:1 ratio. The model was optimized and trained using the Huber loss function, L2 regularization term, and Adam optimizer, with an initial learning rate of 0.0001. The mean absolute error (MAE) and coefficient of determination (R²) were used to measure the model's predictive performance. The focusing strategy is implemented based on the predicted defocus distance. l p Divided into three levels, l p For the range ∈(10,+∞), a large defocusing condition is encountered, and a large step size is used to quickly approach the optimal focus position. l p ∈(1,10] represents the mid-focus area, using a medium step size to balance speed and accuracy. l p ∈(0,1] represents small defocus, which is finely adjusted using micro-steps; the model predicts the defocus distance with a single image acquisition. l p The system then dynamically matches the step size and moves directly to the target area. It then uses the Brenner function to verify the focus quality. If the quality is not up to standard, it triggers a second small defocus adjustment.

3. The deep learning-based visual inspection and grading system for concrete surface roughening quality according to claim 1, characterized in that: The real-time optimization includes using the multi-scale Retinex algorithm to decompose the incident and reflected light components, and using multi-scale spatial convolution of Gaussian filter kernels to eliminate illumination unevenness.

4. The deep learning-based visual inspection and grading system for concrete surface roughening quality according to claim 1, characterized in that, The quality level evaluation includes: （3） （4） （5） （6） （7） （8） In the formula, H - Maximum vertical height difference between adjacent peaks and troughs in a single scan region, mm S - Maximum peak and maximum trough slope in a single scan region N -Number of wave crests per unit area z p,i -No. i The height of each peak, in mm. z v,i -and the i The height of the troughs adjacent to each peak, in mm. x mp , y mp , z mp - Coordinates of the maximum peak, ( x mv , y mv , z mv - Coordinates of the maximum trough n - Number of peaks in a single scan region s - The area of ​​the region scanned in a single scan.

5. The deep learning-based visual inspection and grading system for concrete surface roughening quality according to claim 1, characterized in that: The original image data was stitched together using an improved SIFT feature matching algorithm combined with RANSAC, with a matching point threshold of ≥50 pairs and a homography matrix reprojection error of ≤1.5 pixels.

Images