A dense circuit pattern feature matching method based on DINOv2
By combining the SuperPoint and DINOv2 models, key points and region features of circuit patterns are extracted. Clustering and attention mechanisms are used for image alignment and matching, which solves the problem of low efficiency in traditional methods for circuit pattern matching and achieves efficient and accurate image matching.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHWEST INST OF ELECTRONIC EQUIP TECH (SECOND RES INST OF CHINA ELECTRONICS TECH GRP CORP)
- Filing Date
- 2026-02-28
- Publication Date
- 2026-06-02
AI Technical Summary
Traditional image matching methods are inefficient and inaccurate when processing circuit patterns, especially in ultra-high resolution images, where they struggle to effectively extract and process complex details and noise.
The SuperPoint algorithm is used to extract key point features, combined with the DINOv2 model for region feature matching, and K-means clustering and attention mechanism are used for image alignment and matching. Self-attention and cross-attention are used to learn the process of acquiring corresponding points, so as to achieve accurate image matching.
It improves the accuracy and efficiency of image matching, has good generalization ability and adaptability, and can effectively capture the minute details and structural features of circuit graphics, reducing errors caused by rotation, translation and scaling.
Smart Images
Figure CN122135387A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image processing technology, specifically a dense circuit pattern feature matching method based on DINOv2. Background Technology
[0002] In industrial production, circuit pattern feature matching is a crucial quality control technology, aiming to improve inspection accuracy and efficiency through high-precision image processing methods. Semiconductor chips require precise positioning and inspection during production to ensure that each chip meets quality and functional standards. However, traditional image matching methods suffer from low matching efficiency and accuracy when processing such images due to the large number of feature points and repetitive regions. While deep learning methods have improved matching accuracy to some extent, their computational demands on ultra-high-resolution image processing are extremely high, leading to inefficiency. Therefore, developing a new image matching method to solve the problem of matching ultra-high-resolution images with a large number of repetitive regions has become an urgent issue in industrial production.
[0003] In typical industrial production, circuit diagram data often contains many similar elements, making traditional feature matching methods such as SIFT and ORB ineffective. Extracting effective features from large images poses a challenge for both traditional algorithms and some deep learning models. Large images may contain a wealth of complex details and noise, requiring efficient algorithms to extract and process these features. Summary of the Invention
[0004] To address the problem that traditional feature matching methods are ineffective for matching circuit pattern features, especially for ultra-high resolution images, this invention proposes a dense circuit pattern feature matching method based on DINOv2, aiming to achieve more accurate and efficient image matching.
[0005] This invention is achieved using the following technical solution: a dense circuit pattern feature matching method based on DINOv2, comprising the following steps:
[0006] S1. Image preprocessing;
[0007] S2. Image Coarse Alignment: After reducing the resolution of the two matched images to a level suitable for the deep learning model, the SuperPoint algorithm is used to extract key point features for each image, and the image is divided into several regions. In each region, the DINOv2 model is used to extract region features. Based on each key point feature, the region features of the corresponding key point are superimposed to form candidate point features. Each image is then clustered using the K-means clustering algorithm to obtain K cluster centers for all candidate points. The two matched images are coarsely aligned based on nearest neighbors using the cluster centers. If the distance between the center points of two cluster centers in the image is less than a set threshold, then there is a matching relationship between the two cluster centers, and the region correspondence in the two images is obtained.
[0008] S3, Image Fine Matching:
[0009] (1) Candidate point selection
[0010] For cluster centers with matching relationships in coarse matching, retain all candidate points contained therein;
[0011] (2) Constructing the association graph
[0012] The candidate points retained in the two matched images are used as graph nodes. The graph nodes in the same image are connected to form an inner graph, and the graph nodes in the two matched images are connected to form an outer graph, thus establishing a connection between the two matched images.
[0013] (3) Attention-based acquisition of corresponding points
[0014] Learning stage:
[0015] Based on two matched images, for the connections in the inner and outer graphs, self-attention is used to propagate the connection in the inner graph, and cross-attention is used to propagate the connection in the outer graph, thereby learning the relevant parameters of the process of obtaining corresponding points in the two matched images.
[0016] Reasoning stage:
[0017] For two images to be matched, based on the constructed association graph, the corresponding points of the two images are obtained by using the relevant parameters of the learned corresponding point acquisition process;
[0018] (4) Matching
[0019] Based on the corresponding points of the two images to be matched, the relevant parameters for image matching are obtained using coordinate formulas, thereby achieving image matching.
[0020] The aforementioned dense circuit pattern feature matching method based on DINOv2 performs noise reduction and contrast enhancement on two matched images during image preprocessing to improve image quality.
[0021] The above-mentioned dense circuit pattern feature matching method based on DINOv2, in the image coarse alignment, assumes... and Two matched images and The extracted set of key points, for each key point and Features, superimposed with region features extracted using the DINOv2 model. and Forming candidate point features and .
[0022] This method uses the DINOv2 model to extract dense features from images and combines it with the SuperPoint algorithm for keypoint detection and description. The SuperPoint algorithm can identify regions with unique appearances and structures in images, which typically exhibit good matching properties between images, detecting keypoints by learning salient features in the image. Furthermore, by introducing the advanced DINOv2 model for feature extraction and matching, it overcomes the limitations of traditional matching methods. The improved model can generate a comprehensive feature representation for each keypoint, which includes not only the local appearance information of the keypoint but also the global contextual information of the image, thereby improving the accuracy and generalization ability of matching and effectively capturing subtle details and structural features in circuit graphics images.
[0023] Compared with the prior art, the advantages of this invention are:
[0024] (1) It has strong generalization ability, integrating the knowledge of the DINOv2 model to guide the feature matching process, enabling it to perform well in unseen training domains. This differs from many existing techniques that focus on specific domains or datasets. This generalization ability allows the invention to maintain high performance across a variety of visual tasks and different image domains, especially in new domains.
[0025] (2) An attention mechanism guided by keypoint location is introduced, which separates spatial and appearance information, avoiding over-reliance on positional patterns in the training data. This design improves the model's adaptability to different image distributions and enhances its performance in diverse scenarios.
[0026] (3) A coarse alignment was designed, which roughly aligns the image and divides it into k pairs of paired regions. This method avoids the problem of slow efficiency caused by excessively large image resolution, and at the same time reduces the errors caused by image rotation, translation and scaling. Attached Figure Description
[0027] Figure 1 This is an overall system flowchart of the method of the present invention, used for dense circuit graphic feature matching. Detailed Implementation
[0028] A dense circuit image feature matching method based on DINOv2 is proposed. This method first uses the SuperPoint algorithm and the DINOv2 model to extract image features, and then uses k-means clustering to segment the image into single, non-repeating small regions. This method effectively reduces matching errors caused by excessively high image resolution, thereby significantly improving the accuracy and efficiency of image matching. Specifically, it includes the following steps:
[0029] 1. Image preprocessing:
[0030] The two matched images are then denoised and their contrast enhanced to improve image quality. This step is fundamental to ensuring the smooth progress of subsequent processing steps, helping to reduce noise interference and highlight image features.
[0031] 2. Coarse image alignment:
[0032] After reducing the resolution of the two matched images to a level suitable for the deep learning model, the SuperPoint algorithm is used to extract key point features for each image, and the image is divided into several regions. In each region, the DINOv2 model is used to extract region features. Based on each key point feature, the region features of the corresponding key point are superimposed to form candidate point features. Each image is then clustered using the K-means clustering algorithm to obtain K cluster centers for all candidate points. The two matched images are coarsely aligned based on nearest neighbors using the cluster centers. If the distance between the center points of two cluster centers in the image is less than a set threshold, then there is a matching relationship between the two cluster centers, thus obtaining the region correspondence in the two images.
[0033] 3. Fine-grained image matching
[0034] (1) Candidate point selection
[0035] For cluster centers with matching relationships in coarse matching, retain all candidate points contained therein.
[0036] (2) Constructing the association graph
[0037] Candidate points retained in two matched images are used as graph nodes. Graph nodes in the same image are connected to form an inner graph, and graph nodes in two matched images are connected to form an outer graph, thus establishing a connection between the two matched images.
[0038] (3) Attention-based acquisition of corresponding points
[0039] Learning stage:
[0040] Based on two matched images, for connections in the inner and outer graphs, self-attention is used in the inner graph and cross-attention is used in the outer graph to learn the relevant parameters of the process of obtaining corresponding points in the two matched images.
[0041] Reasoning stage:
[0042] For two images to be matched, based on the association graph constructed after image preprocessing and coarse alignment, the corresponding points of the two images to be matched are obtained using the relevant parameters of the learned corresponding point acquisition process.
[0043] (4) Matching
[0044] Based on the corresponding points of the two images to be matched, the relevant parameters for image matching are obtained using coordinate formulas, thereby achieving image matching.
[0045] In image coarse alignment, set and Two matched images and The extracted set of key points, for each key point and Features, superimposed with region features extracted using the DINOv2 model. and Forming candidate point features and .
Claims
1. A dense circuit pattern feature matching method based on DINOv2, characterized in that: Includes the following steps: S1. Image preprocessing; S2. Image Coarse Alignment: After reducing the resolution of the two matched images to a level suitable for the deep learning model, the SuperPoint algorithm is used to extract key point features for each image, and the image is divided into several regions. In each region, the DINOv2 model is used to extract region features. Based on each key point feature, the region features of the corresponding key point are superimposed to form candidate point features. Each image is then clustered using the K-means clustering algorithm to obtain K cluster centers for all candidate points. The two matched images are coarsely aligned based on nearest neighbors using the cluster centers. If the distance between the center points of two cluster centers in the image is less than a set threshold, then there is a matching relationship between the two cluster centers, and the region correspondence in the two images is obtained. S3, Image Fine Matching: (1) Candidate point selection For cluster centers with matching relationships in coarse matching, retain all candidate points contained therein; (2) Constructing the association graph The candidate points retained in the two matched images are used as graph nodes. The graph nodes in the same image are connected to form an inner graph, and the graph nodes in the two matched images are connected to form an outer graph, thus establishing a connection between the two matched images. (3) Attention-based acquisition of corresponding points Learning stage: Based on two matched images, for the connections in the inner and outer graphs, self-attention is used to propagate in the inner graph, and cross-attention is used to propagate in the outer graph, learning the relevant parameters of the process of obtaining corresponding points in the two matched images. Reasoning stage: For two images to be matched, based on the constructed association graph, the corresponding points of the two images are obtained by using the relevant parameters of the learned corresponding point acquisition process; (4) Matching Based on the corresponding points of the two images to be matched, the relevant parameters for image matching are obtained using coordinate formulas, thereby achieving image matching.
2. The dense circuit pattern feature matching method based on DINOv2 according to claim 1, characterized in that: In image preprocessing, noise reduction and contrast enhancement are performed on the two matched images to improve image quality.
3. A dense circuit pattern feature matching method based on DINOv2 according to claim 1 or 2, characterized in that: In image coarse alignment, set and Two matched images and The extracted set of key points, for each key point and Features, superimposed with region features extracted using the DINOv2 model. and Forming candidate point features and .