Image recognition method based on land features

Abstract

The invention provides an image recognition method based on land features, and the method comprises the steps: S11, carrying out the positioning processing of a to-be-recognized part of a land feature image, obtaining the corresponding coordinates of the to-be-recognized part of the land feature image, projecting the boundary line of a vector pattern spot to the to-be-recognized part of the land feature image, and obtaining a proof image; s12, segmenting the proof image through an SAM segmentation algorithm to obtain a to-be-identified area; s13, performing low-element identification on the to-be-identified region to obtain a corresponding land type; and S14, after the identified land type is matched with the change investigation land type, classifying the land feature image. When the change investigation land feature image is checked, automatic classification can be carried out according to the proof photo and the land feature image, and an auxiliary effect on change investigation is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of image analysis, and in particular to an image recognition method based on land features. Background Art

[0002] The land change survey is an update survey on changes in the current land use status. It refers to the use of aerospace remote sensing, satellite navigation positioning, geographic information, artificial intelligence recognition and other technical means to ensure the timeliness and accuracy of the land survey results based on the unified point survey results of the second and third national land surveys and the land change survey results of the previous year. The change maps are extracted through the latest remote sensing images, and the county-level natural resources management departments conduct efficient on-site investigations and evidence, and timely update the relevant information of the maps to fully grasp the key information such as the spatial distribution, quantity, quality, land use status, and change trends of natural resources, objectively analyze the basic conditions of natural resources and the degree of protection, development and utilization, and give full play to the basic, timely and leading role of natural resource surveys and monitoring to meet the current needs of natural resource management.

[0003] During land change surveys, when verifying land feature images from these changes, automatic classification based on supporting photographs and land feature images is necessary. However, current technical approaches still face numerous challenges. While deep learning methods have demonstrated excellent performance in image segmentation and classification, the remote sensing images used in land change surveys contain a variety of land features, such as buildings, vegetation, and roads. These features may occlude each other and experience significant scale differences within the image. This makes it difficult for deep learning methods to effectively distinguish local, refined targets in land use classification. For example, in remote sensing image analysis, if unsegmented images containing buildings, vegetation, and roads are directly identified, the model may misclassify them due to texture clutter. Therefore, when implementing automatic classification of land feature images, further algorithm optimization is needed to improve the model's adaptability to complex scenarios and classification accuracy. Summary of the Invention

[0004] The purpose of the present invention is to provide an image recognition method based on land features, which can automatically classify land features according to evidence photos and land feature images when verifying land feature images for change investigation.

[0005] To achieve the above-mentioned object of the invention, the present invention provides an image recognition method based on land features, the method comprising:

[0006] S11, performing positioning processing on the portion of the land feature image to be identified, obtaining corresponding coordinates of the portion of the land feature image to be identified, projecting the boundary line of the vector map onto the portion of the land feature image to be identified, and obtaining an evidence image;

[0007] S12, segmenting the evidence image using the SAM segmentation algorithm to obtain the area to be identified;

[0008] S13, performing low-level element identification on the area to be identified to obtain the corresponding land type;

[0009] S14. After matching the identified land type with the changed survey land type, the land feature image is classified.

[0010] Furthermore, in step S11, positioning processing is performed on the portion of the land feature image to be identified, including but not limited to obtaining the current latitude and longitude, altitude, pitch angle, horizontal field of view angle, and vertical field of view angle of the shooting device.

[0011] Furthermore, in step S11, the boundary points of the vector map patches are projected onto the portion of the land feature image to be identified, which specifically includes:

[0012] S21, establishing a bidirectional mapping function between the coordinates of the vector map spots and the corresponding coordinates of the portion to be identified in the land feature image;

[0013] S22. Projecting the boundary points of the vector patch onto the portion of the land feature image to be identified through overlay analysis and a bidirectional mapping function to obtain the patch boundary within the vector range;

[0014] S23. Eliminate the land feature image content outside the map block boundary to obtain an evidence image for the change investigation vector map block.

[0015] Furthermore, in step S12, the evidence image is segmented using the SAM segmentation algorithm, specifically including:

[0016] S31, generating a one-time image embedding by passing the evidence image through an image encoder;

[0017] S32, encoding preset background points, masks, bounding boxes or text into embedding vectors in real time by prompting the encoder;

[0018] S33. Based on the one-time image embedding and the embedding vector, the mask decoder predicts the segmentation mask to output a token, and the one-time image embedding, the embedding vector and the output token are mapped to the mask to generate a segmentation mask map to obtain the area to be identified.

[0019] Furthermore, when segmenting the image patch boundaries using the SAM segmentation algorithm, the CUDA model is used to schedule and allocate CPU and GPU resources, and parallel computing is achieved by configuring kernel functions and setting thread scheduling.

[0020] Furthermore, when segmenting the patch boundaries using the SAM segmentation algorithm, parameters are tuned, including but not limited to filter thresholds and image sampling points, and variable segmentation results are generated based on the tuning parameters.

[0021] Furthermore, after segmenting the patch boundaries using the SAM segmentation algorithm, the processed image is saved as a single element.

[0022] Furthermore, in step S13, the area to be identified is subjected to low-element identification, specifically including:

[0023] S41, identifying and obtaining the outlines of each type of land in the area to be identified, and determining the element information corresponding to the outlines of each type of land based on the embedded vector;

[0024] S42, traverse all data points in the area to be identified, and assign the same color scale to data points at the same level according to the element information;

[0025] S43. Construct a land classification recognition model based on a neural network, input the area to be recognized processed in steps S41 and S42 into the land classification recognition model, and obtain the corresponding land classification type.

[0026] Furthermore, in step S14, the identified land type is matched with the changed survey land type, specifically including:

[0027] Establish a database based on the land classification of the change survey, establish a table mapping relationship based on the land type, and obtain a mapping table;

[0028] The identified land type is matched with the mapping table to obtain the result, and labels are formulated based on the result to automatically classify the land feature images.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The land feature-based image recognition method provided by the present invention, compared to traditional photo recognition and extraction, incorporates photo positioning technology. By fusing the portion of the land feature image to be identified with the vector map patch to form an evidence image, the required specific portion of the land feature image is located. The evidence image is then precisely segmented using the SAM segmentation algorithm. The segmented area to be identified undergoes low-level precise identification, and the resulting results are matched with the land classification for the change survey, before the land feature image is classified. When verifying the land feature image for the change survey, automatic classification can be performed based on the evidence photo and the land feature image, assisting the change survey. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without any creative work.

[0032] Figure 1 A schematic flow chart of a land feature-based image recognition method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The present invention will be further described in detail below with reference to the accompanying drawings and examples. It will be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only illustrate portions relevant to the present invention, not all structures.

[0034] Reference Figure 1 This embodiment provides an image recognition method based on land features, the method comprising:

[0035] S11. Position the portion of the land feature image to be identified, obtain the corresponding coordinates of the portion, project the boundary line of the vector map onto the portion to be identified, and obtain an evidence image. Positioning the portion of the land feature image to be identified includes, but is not limited to, obtaining the current latitude and longitude, altitude, pitch angle, horizontal field of view, and vertical field of view of the camera.

[0036] In this embodiment, the principle of generating evidence photos is based on a calculation algorithm of the mapping relationship between geographic coordinates and image pixels to realize the range marking of vector map spots on the image: information such as the current latitude and longitude, altitude, pitch angle, horizontal field of view angle, and vertical field of view angle of the shooting device is required.

[0037] In step S11, the boundary points of the vector map are projected onto the part of the land feature image to be identified, which specifically includes:

[0038] S21. Establish a bidirectional mapping function between the coordinates of the vector map spots and the corresponding coordinates of the part to be identified in the land feature image.

[0039] S22. Project the boundary points of the vector patch onto the part of the land feature image to be identified through overlay analysis and a bidirectional mapping function to obtain the boundary lines of the patch within the vector range.

[0040] S23. Eliminate the land feature image content outside the map block boundary to obtain an evidence image for the change investigation vector map block.

[0041] In this embodiment, the camera projection model can mathematically realize the mapping relationship between three-dimensional points in the real world and pixel coordinates on the land feature image. By projecting the boundary points of the vector patch onto the land feature image, the patch boundary can be formed on the land feature image, thereby accurately extracting the land feature information within the vector patch. The generated evidence image is for the change investigation vector patch and can intuitively display the changes in land features within the vector patch. This step can combine complex, multi-source land feature images with actual vector patches to obtain a more accurate evidence image range, providing an accurate data source for subsequent image segmentation.

[0042] S12. Segment the evidence image using the SAM segmentation algorithm to obtain the area to be identified. Segmenting the evidence image using the SAM segmentation algorithm specifically includes:

[0043] S31. Generate a one-time image embedding by passing the evidence image through an image encoder.

[0044] S32, encoding preset background points, masks, bounding boxes or text into embedding vectors in real time by prompting the encoder.

[0045] S33. Based on the one-time image embedding and the embedding vector, the mask decoder predicts the segmentation mask to output a token, and the one-time image embedding, the embedding vector and the output token are mapped to the mask to generate a segmentation mask map to obtain the area to be identified.

[0046] In this embodiment, image segmentation mainly uses the SAM feature segmentation algorithm as the core basic element. The SAM algorithm is used to segment the evidence image. The image encoder uses a convolutional neural network (CNN) in deep learning, such as ResNet, EfficientNet or Vision Transformer (ViT), and generates a one-time image embedding through operations such as convolutional layers, pooling layers and fully connected layers; the prompt encoder encodes complex prompt information, such as background points, masks, bounding boxes or text, into embedding vectors in real time, thereby realizing automated image segmentation and recognition. For example, in land surveys, prompt information can be dynamically adjusted according to different survey targets, such as specific types of land features or patches, thereby improving recognition accuracy. Based on the one-time image embedding and embedding vector, the mask decoder accurately predicts the segmentation mask, effectively distinguishes the features inside and outside the patch, avoids misjudgment and missed judgment, and improves the accuracy of land feature recognition.

[0047] When segmenting the image patch boundaries using the SAM segmentation algorithm, the CUDA model is used to schedule and allocate CPU and GPU resources, and parallel computing is achieved by configuring kernel functions and setting thread scheduling.

[0048] In this embodiment, CUDA (Compute Unified Device Architecture) is a parallel computing platform and programming model launched by NVIDIA. It significantly reduces computing time by assigning computationally intensive tasks to the GPU, such as forward propagation and backpropagation of deep learning models. In CUDA, kernel functions are functions that run on the GPU, and thread scheduling is achieved by configuring thread blocks and thread grids to achieve parallel computing. The GPU has a large number of CUDA cores and can process multiple threads at the same time, thereby achieving parallel computing. For convolution operations and matrix operations in the SAM segmentation algorithm, parallel processing can greatly improve computing efficiency. By reasonably allocating tasks between the CPU and GPU, for example, assigning tasks such as data preprocessing and I / O operations to the CPU, and assigning the computing tasks of the deep learning model to the GPU, efficient use of resources can be achieved.

[0049] Land surveys and segmentation tasks often require processing large numbers of high-resolution images. CUDA acceleration allows for rapid processing of these images, meeting real-time and large-scale processing requirements. For applications requiring real-time feedback, such as dynamic monitoring of land use changes, CUDA acceleration ensures that segmentation tasks are completed in a short period of time, providing timely analysis results. For the SAM model, performance can be optimized by adjusting the model's parallelization strategy. For example, using PyTorch's DataParallel or DistributedDataParallel modules to implement multi-GPU parallel computing, the preprocessed images are fed into the SAM model to perform the segmentation task. Since the model is already loaded onto the GPU, the computational process is automatically accelerated using CUDA.

[0050] When segmenting the image boundary using the SAM segmentation algorithm, parameters are tuned, including but not limited to the filter threshold and image sampling points, to generate variable segmentation results based on the debugging parameters. After segmenting the image boundary using the SAM segmentation algorithm, the processed image is saved as a single element.

[0051] In this embodiment, adjusting the filter threshold can remove noise or smooth segmentation boundaries, thereby improving the accuracy of the segmentation results. Adjusting the number and distribution of sampling points can optimize the segmentation algorithm's ability to capture detail, further enhancing the segmentation effect. Saving processed images as individual elements facilitates subsequent analysis, management, and application. For example, in a land survey, the segmentation results for each patch can be saved as a separate file for further statistics, monitoring, and display.

[0052] S13. Perform low-element identification on the area to be identified to obtain the corresponding land type.

[0053] Identify the low-level elements in the area to be identified, including:

[0054] S41, identifying and obtaining the outlines of each type of land in the area to be identified, and determining the element information corresponding to the outlines of each type of land based on the embedded vector;

[0055] S42, traverse all data points in the area to be identified, and assign the same color scale to data points at the same level according to the element information;

[0056] S43. Construct a land classification recognition model based on a neural network, input the area to be recognized processed in steps S41 and S42 into the land classification recognition model, and obtain the corresponding land classification type.

[0057] In this embodiment, low-element recognition focuses on the basic features and outlines of land classes. It is suitable for land use classification in complex scenarios and can effectively handle situations where land class boundaries are blurred or overlapping. By creating a mask or directly modifying image pixels, the same color scale is assigned to data points at the same level, enhancing the visual effect of the image and facilitating subsequent analysis and processing. By identifying the outline of the land class and combining the embedded vector to determine the element information, the shape and characteristics of the land class are more accurately described, thereby improving recognition accuracy. A land class recognition model is constructed using a deep learning framework based on a neural network, the model is trained using labeled data, and reasoning is performed on the processed area to be identified to obtain the land class type.

[0058] S14. After matching the identified land type with the changed survey land type, the land feature image is classified. Matching the identified land type with the changed survey land type specifically includes:

[0059] Establish a database based on the land classification of the change survey, establish a table mapping relationship based on the land type, and obtain a mapping table;

[0060] The identified land type is matched with the mapping table to obtain the result, and labels are formulated based on the result to automatically classify the land feature images.

[0061] In this embodiment, in order to support the change survey business, a database is constructed according to the change survey land classification, a table mapping relationship is established according to the land type, and then the results are obtained by matching the identified land type and the mapping table, and labels are formulated according to the results to automatically classify the land feature images.

[0062] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An image recognition method based on land features, characterized in that: The method comprises: S11, performing positioning processing on the portion of the land feature image to be identified, obtaining corresponding coordinates of the portion of the land feature image to be identified, projecting the boundary line of the vector map onto the portion of the land feature image to be identified, and obtaining an evidence image; S12, segmenting the evidence image using the SAM segmentation algorithm to obtain the area to be identified; S13, performing low-level element identification on the area to be identified to obtain the corresponding land type; S14. After matching the identified land type with the changed survey land type, the land feature image is classified.

2. The land feature-based image recognition method according to claim 1, characterized in that: In step S11 , positioning processing is performed on the portion of the land feature image to be identified, including but not limited to obtaining the current latitude and longitude, altitude, pitch angle, horizontal field of view angle, and vertical field of view angle of the shooting device.

3. The land feature-based image recognition method according to claim 1, characterized in that: In step S11, the boundary points of the vector map are projected onto the part of the land feature image to be identified, which specifically includes: S21, establishing a bidirectional mapping function between the coordinates of the vector map spots and the corresponding coordinates of the portion to be identified in the land feature image; S22. Projecting the boundary points of the vector patch onto the portion of the land feature image to be identified through overlay analysis and a bidirectional mapping function to obtain the patch boundary within the vector range; S23. Eliminate the land feature image content outside the map block boundary to obtain an evidence image for the change investigation vector map block.

4. The land feature-based image recognition method according to claim 1, characterized in that: In step S12, the evidence image is segmented using the SAM segmentation algorithm, specifically including: S31, generating a one-time image embedding by passing the evidence image through an image encoder; S32, encoding preset background points, masks, bounding boxes or text into embedding vectors in real time by prompting the encoder; S33. Based on the one-time image embedding and the embedding vector, the mask decoder predicts the segmentation mask to output a token, and the one-time image embedding, the embedding vector and the output token are mapped to the mask to generate a segmentation mask map to obtain the area to be identified.

5. The land feature-based image recognition method according to claim 4, characterized in that: When segmenting the image patch boundaries using the SAM segmentation algorithm, the CUDA model is used to schedule and allocate CPU and GPU resources, and parallel computing is achieved by configuring kernel functions and setting thread scheduling.

6. The land feature-based image recognition method according to claim 4, characterized in that: When segmenting the patch boundary using the SAM segmentation algorithm, parameters are tuned, including but not limited to filter thresholds and image sampling points, and variable segmentation results are generated based on the debugging parameters.

7. The land feature-based image recognition method according to claim 4, characterized in that: After segmenting the image patch boundaries using the SAM segmentation algorithm, the processed image is saved as a single element.

8. The land feature-based image recognition method according to claim 4, characterized in that: In step S13, the low-element recognition is performed on the area to be recognized, which specifically includes: S41, identifying and obtaining the outlines of each type of land in the area to be identified, and determining the element information corresponding to the outlines of each type of land based on the embedded vector; S42, traverse all data points in the area to be identified, and assign the same color scale to data points at the same level according to the element information; S43. Construct a land classification recognition model based on a neural network, input the area to be recognized processed in steps S41 and S42 into the land classification recognition model, and obtain the corresponding land classification type.

9. The land feature-based image recognition method according to claim 1, characterized in that: In step S14, the identified land type is matched with the changed survey land type, specifically including: Establish a database based on the land classification of the change survey, establish a table mapping relationship based on the land type, and obtain a mapping table; The identified land type is matched with the mapping table to obtain the result, and labels are formulated based on the result to automatically classify the land feature images.

Images