An industrial inspection method, system and device based on multi-frame image registration

Abstract

The present invention discloses an industrial detection method, system and device based on multi-frame image registration. The method includes: setting the image acquisition frequency parameter of an industrial camera; acquiring a series of consecutive multi-frame images of a product to be detected; preprocessing the multi-frame images; performing rough registration on the preprocessed multi-frame images; respectively estimating the pixel-level optical flow motion field by two methods of direct estimation and indirect estimation; fusing the pixel-level optical flow fields of direct estimation and indirect estimation; aligning the multi-frame images to a reference frame according to the fine registration result; inputting the aligned multi-frame images into a multi-frame super-resolution network for feature extraction and fusion, and outputting a super-resolution reconstructed image; and inputting the super-resolution reconstructed image into a detection network to output the detection result of the product to be detected. Through multi-frame image acquisition, preprocessing, multi-frame registration, super-resolution reconstruction and detection, the dependence on high-precision detection equipment and high environmental requirements is effectively reduced, and efficient and accurate industrial detection is achieved while reducing the cost of hardware equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of optical image processing, and more specifically, to an industrial detection method, system and device based on multi-frame image registration. Background Art

[0002] Currently, with the gradual improvement of industrial automation, some detection systems based on traditional image processing algorithms have begun to be applied. These early image processing detection technologies mainly use methods such as simple threshold segmentation, edge detection, and morphological operations to identify features and defects in product images. For example, in the production of electronic components, for the welding quality detection of electronic components on a circuit board, the solder joint area is separated from the background by threshold segmentation, and then the shape and size of the solder joint are judged by edge detection to see if they meet the standards. However, when faced with complex industrial environments and diverse detection requirements, these traditional image processing methods have shown obvious limitations:

[0003] On the one hand, their adaptability to images is poor. Images in actual industrial production are often affected by factors such as uneven lighting conditions, object surface reflection, background noise interference, and changes in the shape and material of the product itself, resulting in difficulties for traditional algorithms to accurately extract stable and reliable feature information. For example, in the metal processing industry, for the detection of tiny scratches and cracks on the metal surface, due to the reflective characteristics of the metal surface, traditional edge detection algorithms may misjudge the reflective area as a defect edge, resulting in a large number of false detection results; in food packaging and printing detection, due to the diversity of the texture and color of the packaging material, simple threshold segmentation methods cannot effectively distinguish normal printed patterns from defective parts, causing a serious decline in detection accuracy.

[0004] On the other hand, the intelligence level of traditional image processing algorithms is relatively low, and it is difficult to handle complex image patterns and changing detection tasks. During the industrial production process, the types and specifications of products are constantly updated, and the detection requirements are becoming increasingly complex and diverse. It is required that the detection system can automatically identify different types of defects and accurately evaluate and classify the severity of the defects. However, traditional algorithms usually can only detect specific, pre-set defect patterns and lack the ability of automatic learning and recognition of unknown or newly emerging defects, unable to meet the needs of intelligent detection in modern industrial production.

[0005] In addition, with the development of industrial production towards high precision and microscale, the requirements for the accuracy and resolution of image detection are also getting higher and higher. For example, in the field of semiconductor chip manufacturing, the circuit line width on the chip has reached the nanometer level, and high-precision image analysis capabilities at the sub-micron or even nanometer level are required for defect detection on the chip surface. When traditional image processing technologies process high-resolution images, the computational complexity increases sharply, and the processing speed drops significantly, making it difficult to meet the requirements of real-time online detection.

[0006] In recent years, the rise of deep learning technology has brought new development opportunities to industrial inspection of image processing. Deep learning algorithms, such as convolutional neural networks (CNNs), can automatically learn complex feature representations from a large amount of image data by constructing multi-layer neural network structures, and have achieved remarkable results in tasks such as image classification, object detection, and semantic segmentation. Some advanced industrial inspection systems have begun to try to introduce deep learning technology to improve inspection performance. For example, in the surface defect detection of photovoltaic cells, using a deep learning model to train and identify the images of the cells can effectively detect various types of tiny defects and improve the accuracy and stability of inspection to a certain extent.

[0007] However, the application of deep learning technology in industrial inspection is still in the exploratory and development stage and faces some problems that need to be solved urgently: First, deep learning models are highly dependent on a large amount of labeled data. In the industrial field, obtaining high-quality and large-scale labeled data often requires a large amount of human, material, and time costs, and the quality of data labeling also has an important impact on the performance of the model. Second, deep learning models have poor interpretability and it is difficult to intuitively understand how the model makes decisions, which brings certain troubles in aspects such as quality traceability and fault diagnosis in industrial production. In addition, deep learning models have high computational resource requirements and need to be equipped with high-performance hardware devices, such as GPU clusters, which increases the equipment procurement and maintenance costs of enterprises and limits their application in some small enterprises and environments with limited resources.

[0008] In summary, although the current industrial inspection technology of image processing has made certain progress, there are still many problems and challenges in terms of accuracy, adaptability, intelligence level, and cost-effectiveness. Summary of the Invention

[0009] The purpose of the present invention is to overcome the above-mentioned deficiencies of the prior art and provide an industrial inspection method, system, and device based on multi-frame image registration. Through multi-frame image acquisition, preprocessing, multi-frame registration, super-resolution reconstruction, and detection, it effectively reduces the dependence on high-precision inspection equipment and high environmental requirements, realizes efficient and accurate industrial inspection while reducing the hardware equipment cost, and meets the urgent needs of modern industrial production for cost-controllable, high-quality, high-efficiency, and intelligent inspection.

[0010] To achieve the above purpose, the present invention adopts the following technical solutions:

[0011] An industrial inspection method based on multi-frame image registration, the method includes:

[0012] Step S1, setting the image acquisition frequency parameter of the industrial camera according to the number of frames of multi-frame images to be collected;

[0013] Step S2, collect consecutive multiple frames of images of the product to be detected through an industrial camera;

[0014] Step S3, preprocess the obtained multiple frames of images, and the preprocessing includes removing the first and last frames, removing abnormal frames, and image denoising;

[0015] Step S4, perform rough registration on the preprocessed multiple frames of images to obtain the global transformation from the image to be registered to the reference image;

[0016] Step S5, on the basis of rough registration, estimate the pixel-level optical flow motion field by direct estimation and indirect estimation respectively;

[0017] Step S6, fuse the pixel-level optical flow fields estimated by direct estimation and indirect estimation to obtain the fine registration result from the frame to be registered to the reference frame;

[0018] Step S7, align the multiple frames of images to the reference frame respectively according to the fine registration result to obtain the aligned multiple frames of images;

[0019] Step S8, input the aligned multiple frames of images into a multi-frame super-resolution network for feature extraction and fusion, and output the super-resolution reconstructed image;

[0020] Step S9, input the super-resolution reconstructed image into the detection network and output the detection result of the product to be detected.

[0021] Furthermore, in step S1, the formula for calculating the image acquisition frequency of the industrial camera is as follows:

[0022]

[0023] In the formula, N is the number of multiple frames of images to be collected; is the time of the last frame of image when the acquisition ends; is the time of the first frame of image when the acquisition is triggered.

[0024] Furthermore, in step S3, the abnormal frame removal includes the following steps:

[0025] Step S301, calculate the grayscale histogram of each frame of image ;

[0026]

[0027] In the formula, is the pixel grayscale value of the i-th frame of image at the position k is the grayscale level, is the Dirac function;

[0028] Step S302, calculate the average grayscale histogram of multiple frames of images as a benchmark ;

[0029]

[0030] Step S303: Calculate the grayscale histogram of each frame of image and the reference difference ;

[0031] Step S304: Calculate the deviation between the difference of each frame of image and the overall distribution , when , determine it as an abnormal frame and eliminate it. The deviation calculation formula is as follows:

[0032]

[0033] In the formula, and are the mean and standard deviation of the histogram difference respectively.

[0034] Furthermore, in Step S4, the rough image registration includes the following steps:

[0035] Step S401: Select one frame from multiple frames of images as the reference frame , and the remaining frames as the frames to be registered ;

[0036] Step S402: For the reference frame and the frames to be registered extract image key points and descriptors respectively;

[0037] Step S403: Match the descriptors of the reference frame and the frames to be registered through the nearest neighbor algorithm to obtain the matching point pairs between the reference frame and the frames to be registered ;

[0038] Step S404: Use the RANSAC algorithm to eliminate the mismatched points;

[0039] Step S405: Calculate the global transformation matrix T according to the inlier set after eliminating the mismatched points;

[0040] Step S406: Transform the frames to be registered through the transformation matrix T to obtain the roughly registered frames .

[0041] Further, in step S5, in the direct estimation method, the pixel-level optical flow motion field of the frame to be registered relative to the reference frame is calculated. In the indirect estimation method, after calculating the inter-frame pixel-level optical flow motion fields of the front and rear frames within the interval between the frame to be registered and the reference frame, the pixel-level optical flow motion field of the frame to be registered relative to the reference frame is obtained by superposition.

[0042] Further, in step S5, the implementation method of directly estimating the pixel-level optical flow motion field is as follows:

[0043] Use the Lucas-Kanade optical flow method to calculate the coarsely registered frame to the reference frame of the pixel-level optical flow field

[0044] .

[0045] Further, in step S5, the implementation method of indirectly estimating the pixel-level optical flow motion field is as follows:

[0046] Step S501, for the intermediate frame between the frame to be registered and the reference frame , , obtain the coarsely registered frame from the intermediate frame to the reference frame through step S4;

[0047] Step S502, for the previous frame of the intermediate frame , use the Lucas-Kanade optical flow method to calculate the coarsely registered frame to the coarsely registered frame of the pixel-level optical flow field

[0048] .

[0049] Further, in step S6, the implementation method of fusing the directly estimated and indirectly estimated pixel-level optical flow fields is as follows:

[0050]

[0051] In the formula, is the fine registration result of the frame to be registered to the reference frame .

[0052] The present invention also provides an industrial detection system based on multi-frame image registration. The system includes:

[0053] A multi-frame image acquisition unit for photographing the product to be detected and acquiring a continuous multi-frame image;

[0054] A preprocessing unit for preprocessing multiple frames of images by removing the first and last frames, removing abnormal frames, and denoising the images;

[0055] A multi-frame registration unit for aligning the preprocessed multiple frames of images to a reference frame to obtain the aligned multiple frames of images;

[0056] A super-resolution reconstruction unit for extracting and fusing features from the aligned multiple frames of images to obtain a super-resolution reconstructed image;

[0057] A detection unit for detecting the super-resolution reconstructed image and outputting the detection result of the product to be detected.

[0058] The present invention also provides an industrial detection device based on multi-frame image registration, including:

[0059] A memory for storing computer programs and data;

[0060] A processor for implementing the steps of the industrial detection method based on multi-frame image registration as described above when executing the computer program.

[0061] The beneficial effects of the present invention are:

[0062] The present invention can improve the detection efficiency restricted by low image quality caused by noisy environment, uneven illumination, and limited device performance in industrial detection. By collecting, preprocessing, multi-frame registering, super-resolution reconstructing, and detecting multiple frames of images, it effectively reduces the dependence on high-precision detection devices and high environmental requirements, realizes efficient and accurate industrial detection while reducing the cost of hardware devices, and meets the urgent needs of modern industrial production for cost-control, high-quality, high-efficiency, and intelligent detection. Description of the Drawings

[0063] Figure 1 It is a flowchart of an industrial detection method based on multi-frame image registration in this embodiment;

[0064] Figure 2 It is a structural framework diagram of an industrial detection system based on multi-frame image registration in this embodiment;

[0065] Figure 3 It is a structural framework diagram of an industrial detection device based on multi-frame image registration in this embodiment.

[0066] Reference numerals: multi-frame image acquisition unit 1, preprocessing unit 2, multi-frame registration unit 3, super-resolution reconstruction unit 4, detection unit 5, memory 6, processor 7. Detailed Embodiments

[0067] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0068] Embodiment: An industrial detection method based on multi-frame image registration, as Figure 1 shown, the method includes:

[0069] Step S1, set the image acquisition frequency parameter of the industrial camera according to the number of multi-frames of images to be acquired;

[0070] Among them, the calculation formula of the image acquisition frequency of the industrial camera is as follows:

[0071]

[0072] In the formula, N is the number of multi-frames of images to be acquired, which can generally be set to 5 to 20 frames; is the time when the last frame of the image acquisition ends; is the time when the acquisition of the first frame of the image is triggered.

[0073] Step S2, collect a series of multi-frames of images of the product to be detected through the industrial camera;

[0074] Step S3, preprocess the obtained multi-frames of images, and the preprocessing includes removing the first and last frames, removing abnormal frames, and image denoising;

[0075] Among them, removing the first and last frames means removing the first frame of the image triggered by the industrial camera and the last frame before the end of the acquisition, to avoid the situation where the imaging of the product to be detected is incomplete.

[0076] Removing abnormal frames means removing abnormal images such as black images, overexposure, and underexposure caused by industrial camera failures, changes in the illumination of the imaging environment, etc. The specific steps are as follows:

[0077] Step S301, calculate the gray-level histogram of each frame of the image ;

[0078]

[0079] In the formula, is the pixel gray value of the i-th frame of the image at the position , k is the gray level, is the Dirac function;

[0080] Step S302, calculate the average gray-level histogram of the multi-frames of images as a benchmark ;

[0081]

[0082] Step S303: Calculate the grayscale histogram of each frame of the image. And the reference Difference ; The difference metric adopts one of Euclidean distance, Bhattacharyya distance, and chi-square distance.

[0083] Step S304: Calculate the deviation between the difference of each frame of the image and the overall distribution. When , it is determined as an abnormal frame and excluded. The deviation The calculation formula is as follows:

[0084]

[0085] In the formula, And Are the mean and standard deviation of the histogram difference respectively.

[0086] Image denoising adopts one of median filtering, bilateral filtering, and non-local means filtering.

[0087] Step S4: Coarsely register the preprocessed multi-frame images to obtain the global transformation from the image to be registered to the reference image.

[0088] Among them, the image coarse registration includes the following steps:

[0089] Step S401: Select one frame from the multi-frame images as the reference frame , which can be the middle frame or the frame closest to the average distance; the remaining frames are used as the frames to be registered ;

[0090] Step S402: For the reference frame And the frame to be registered Extract the image key points and descriptors respectively, and one of SIFT, SURF, BRISK, and ORB can be used;

[0091] Step S403: Match the descriptors of the reference frame And the frame to be registered Through the nearest neighbor algorithm (such as k-NN) to obtain the matching point pairs between the reference frame And the frame to be registered ;

[0092] Step S404: Use the RANSAC algorithm to eliminate the mismatched points;

[0093] Step S405: Calculate the global transformation matrix T according to the inlier set after eliminating the mismatched points.

[0094] Step S406, transform the frame to be registered through the transformation matrix T to obtain a roughly registered frame .

[0095] Step S5, on the basis of rough registration, estimate the pixel-level optical flow motion field in two ways: direct estimation and indirect estimation; among them, the direct estimation method calculates the pixel-level optical flow motion field of the frame to be registered relative to the reference frame, and the indirect estimation method calculates the inter-frame pixel-level optical flow motion field of the front and back frames within the interval between the frame to be registered and the reference frame, and then superimposes to obtain the pixel-level optical flow motion field of the frame to be registered relative to the reference frame;

[0096] Furthermore, the implementation method of directly estimating the pixel-level optical flow motion field is:

[0097] Use the Lucas-Kanade optical flow method to calculate the pixel-level optical flow field between the roughly registered frame and the reference frame

[0098] .

[0099] Furthermore, the implementation method of indirectly estimating the pixel-level optical flow motion field is:

[0100] Step S501, for the intermediate frame between the frame to be registered and the reference frame , , obtain the roughly registered frame of the intermediate frame to the reference frame through Step S4;

[0101] Step S502, for the previous frame of the intermediate frame , use the Lucas-Kanade optical flow method to calculate the pixel-level optical flow field from the roughly registered frame of to the roughly registered frame

[0102] .

[0103] Step S6, fuse the pixel-level optical flow fields obtained by direct estimation and indirect estimation to obtain the fine registration result from the frame to be registered to the reference frame;

[0104] Among them, the implementation method of fusing the pixel-level optical flow fields obtained by direct estimation and indirect estimation is:

[0105]

[0106] Wherein, is the to-be-registered frame to the reference frame of the fine registration result.

[0107] Step S7: Align the multiple frames of images to the reference frame respectively according to the fine registration result to obtain the aligned multiple frames of images;

[0108] Step S8: Input the aligned multiple frames of images into a multi-frame super-resolution network for feature extraction and fusion, and output a super-resolution reconstructed image; the multi-frame super-resolution network adopts one of DeepSR, TecoGAN, and MuCAN.

[0109] Step S9: Input the super-resolution reconstructed image into a detection network, and output the detection result of the product to be detected. The detection network adopts one of YOLO and Mask R-CNN.

[0110] The technical solution provided by the embodiment of the present application can improve the detection efficiency restricted by too low image quality caused by noisy environment, uneven illumination, and limited device performance in industrial detection. This method provides an innovative industrial detection solution based on multi-frame image registration through technologies such as multi-frame registration, super-resolution reconstruction, and detection, effectively reducing the dependence on high-precision detection equipment and high environmental requirements, achieving efficient and accurate industrial detection while reducing the cost of hardware equipment, and meeting the urgent needs of modern industrial production for cost-controllable, high-quality, high-efficiency, and intelligent detection.

[0111] See Figure 2 , which is a schematic structural diagram of an industrial detection system based on multi-frame image registration provided by the embodiment of the present application; the system includes a multi-frame image acquisition unit 1, a preprocessing unit 2, a multi-frame registration unit 3, a super-resolution reconstruction unit 4, and a detection unit 5.

[0112] Among them, the multi-frame image acquisition unit 1 is used to photograph the product to be detected and acquire multiple consecutive frames of images;

[0113] The preprocessing unit 2 is used to perform preprocessing on the multiple frames of images, including removing the first and last frames, removing abnormal frames, and image denoising;

[0114] The multi-frame registration unit 3 is used to align the preprocessed multiple frames of images to the reference frame to obtain the aligned multiple frames of images;

[0115] The super-resolution reconstruction unit 4 is used to perform feature extraction and fusion on the aligned multiple frames of images to obtain a super-resolution reconstructed image;

[0116] The detection unit 5 is used to detect the super-resolution reconstructed image and output the detection result of the product to be detected.

[0117] In the technical solution provided by the embodiment of the present application, first, the image acquisition frequency parameter of the industrial camera is set according to the number of multiple frames of images to be acquired; then, the industrial camera is used to acquire multiple consecutive frames of images of the product to be detected; the acquired multiple frames of images are preprocessed, including removing the first and last frames, removing abnormal frames, and image denoising; the preprocessed multiple frames of images are roughly registered to obtain the global transformation from the image to be registered to the reference image; on the basis of rough registration, the pixel-level optical flow motion field is estimated by two methods: direct estimation and indirect estimation; the pixel-level optical flow fields estimated by direct estimation and indirect estimation are fused to obtain the fine registration result from the frame to be registered to the reference frame; then, according to the fine registration result, the multiple frames of images are aligned to the reference frame respectively to obtain the aligned multiple frames of images; the aligned multiple frames of images are input into the multi-frame super-resolution network for feature extraction and fusion, and the super-resolution reconstructed image is output; finally, the super-resolution reconstructed image is input into the detection network, and the detection result of the product to be detected is output.

[0118] See Figure 3 , which is an industrial detection device based on multi-frame image registration provided by the embodiment of the present application; the device includes a memory 6 and a processor 7;

[0119] Among them, the memory 6 is used to store computer programs and data; the processor 7 is used to implement the steps of the above-mentioned industrial detection method based on multi-frame image registration when executing the computer program.

[0120] The above is only the preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions within the idea of the present invention belong to the protection scope of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. An industrial inspection method based on multi-frame image registration, characterized in that, The method includes: Step S1, setting the image acquisition frequency parameter of the industrial camera according to the number of multiple frames of images to be acquired; Step S2, acquiring multiple consecutive frames of images of the product to be detected through the industrial camera; Step S3, preprocessing the acquired multiple frames of images, and the preprocessing includes removing the first and last frames, removing abnormal frames, and image denoising; Step S4, performing rough registration on the preprocessed multiple frames of images to obtain the global transformation from the image to be registered to the reference image; Step S5, on the basis of rough registration, respectively estimating the pixel-level optical flow motion field in two ways: direct estimation and indirect estimation; Step S6, fusing the pixel-level optical flow fields estimated by direct estimation and indirect estimation to obtain the fine registration result from the frame to be registered to the reference frame; Step S7, aligning the multiple frames of images to the reference frame respectively according to the fine registration result to obtain the aligned multiple frames of images; Step S8, inputting the aligned multiple frames of images into a multi-frame super-resolution network for feature extraction and fusion, and outputting a super-resolution reconstructed image; Step S9, inputting the super-resolution reconstructed image into a detection network and outputting the detection result of the product to be detected; In Step S5, in the direct estimation method, the pixel-level optical flow motion field of the frame to be registered relative to the reference frame is calculated, and in the indirect estimation method, after calculating the inter-frame pixel-level optical flow motion field between the frames before and after within the interval of the frame to be registered and the reference frame, the pixel-level optical flow motion field of the frame to be registered relative to the reference frame is obtained by superposition; In Step S6, the implementation method of fusing the pixel-level optical flow fields estimated by direct estimation and indirect estimation is as follows: In the formula, is the fine registration result from the frame to be registered to the reference frame ; is the coarse registration frame of the frame to be registered to the pixel-level optical flow field from the coarse registration frame to the reference frame ; is the pixel-level optical flow field from the coarse registration frame to the coarse registration frame 2. The industrial inspection method based on multi-frame image registration according to claim 1, characterized in that, In Step S1, the calculation formula of the image acquisition frequency of the industrial camera is as follows: Where N is the number of multiple frames of images to be collected; is the time to end the acquisition of the last frame of image; is the time to trigger the acquisition of the first frame of image.

3. The industrial inspection method based on multi-frame image registration according to claim 1, wherein In Step S3, the steps of removing abnormal frames include the following: Step S301, calculate the grayscale histogram of each frame of image ; wherein, is the pixel gray value of the i-th frame image at the position , k is the gray level, is the Dirac function; Step S302, calculate the average grayscale histogram of multiple frames of images as a reference ; Step S303, calculate the grayscale histogram of each frame of image with the reference difference ; Step S304, calculate the deviation of the difference of each frame image from the overall distribution When , it is determined as an abnormal frame and excluded. The deviation The calculation formula is as follows: In the formula, and are the mean and standard deviation of the histogram difference, respectively.

4. The industrial detection method based on multi-frame image registration according to claim 1, wherein In Step S4, the rough registration of the image includes the following steps: Step S401: Select one frame from multiple frames as the reference frame , and use the remaining frames as frames to be registered ; Step S402, for the reference frame and the frame to be registered extract image key points and descriptors respectively; Step S403, match the reference frame by the nearest neighbor algorithm and the frame to be registered descriptors to obtain the matching point pairs between the reference frame and the frame to be registered; Step S404, using the RANSAC algorithm to remove the mismatched points; Step S405, calculating the global transformation matrix T according to the inlier set after removing the mismatched points; Step S406, the frame to be registered is transformed by the transformation matrix T to obtain a roughly registered frame .

5. The industrial detection method based on multi-frame image registration according to claim 1, wherein In Step S5, the implementation method of directly estimating the pixel-level optical flow motion field is as follows: Calculate the pixel-level optical flow field between the rough registration frame and the reference frame using the Lucas-Kanade optical flow method to the reference frame using the Lucas-Kanade optical flow method 。 6. The industrial inspection method based on multi-frame image registration according to claim 1, wherein, In Step S5, the implementation method of indirectly estimating the pixel-level optical flow motion field is as follows: Step S501, for the frame to be registered and the reference frame for the intermediate frame , , obtain the intermediate frame through step S4 to the reference frame of the roughly registered frame ; Step S502, for the intermediate frame of the previous frame , calculate using the Lucas-Kanade optical flow method of the coarsely registered frame to of the coarsely registered frame of the pixel-level optical flow field 。 7. An industrial inspection system based on multi-frame image registration for implementing the method according to claim 1, characterized in that, The system includes: A multi-frame image acquisition unit (1) for photographing the product to be detected and acquiring multiple consecutive frames of images; A preprocessing unit (2) for preprocessing the multiple frames of images, including removing the first and last frames, removing abnormal frames, and image denoising; A multi-frame registration unit (3) for aligning the preprocessed multiple frames of images to the reference frame to obtain the aligned multiple frames of images; A super-resolution reconstruction unit (4) for performing feature extraction and fusion on the aligned multiple frames of images to obtain a super-resolution reconstructed image; A detection unit (5) for detecting the super-resolution reconstructed image and outputting the detection result of the product to be detected.

8. An industrial detection device based on multi-frame image registration, characterized in that, Including: A memory (6) for storing computer programs and data; A processor (7) for implementing the steps of the industrial detection method based on multi-frame image registration as described in any one of claims 1-6 when executing the computer program.

Images