Flame image recognition and feature extraction method and system

Abstract

The invention discloses a flame image recognition and feature extraction method and system, and belongs to the field of engine testing. The method comprises the following steps: performing background noise reduction and boundary enhancement preprocessing on an engine combustion flame image; a preliminary mask is generated by adopting an adaptive threshold method, and fine segmentation is realized in combination with a U-Net network; tracking flame movement by using a background subtraction method and an optical flow method, and analyzing the propagation speed and direction; three-dimensional reconstruction is realized based on MonoDepth2 monocular depth estimation and binocular vision parallax calculation, and a flame space position and an external cube are determined; constructing a convolutional neural network model by taking ResNet-50 as a backbone network, and respectively predicting continuity parameters and discrete parameters through regression branches and classification branches; and fusing the flame area, the speed, the color histogram and the edge features, and finishing final recognition through an SVM (Support Vector Machine) classifier. According to the invention, automatic flame identification, three-dimensional positioning and multi-feature parameter extraction are realized, and manual intervention is reduced.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of engine testing, and particularly relates to a flame image recognition and feature extraction method and system. BACKGROUND

[0002] Engine combustion visualization testing plays an important role in evaluating combustion efficiency, optimizing design and reducing emissions, and can provide detailed data of combustion process, performance and emissions. In the prior art, flame image recognition and feature extraction mainly rely on traditional image processing or single deep learning model. The invention patent CN119091368A identifies the flame area by pre-processing real-time images and deep learning comparison, but this method can only locate the flame position in the two-dimensional plane and cannot predict the three-dimensional spatial coordinates and dynamic propagation characteristics of the flame in the engine combustion process. The invention patent CN118736491A realizes flame recognition based on the comparison of regional temperature and preset ignition threshold, but the engine combustion test needs to extract multi-dimensional parameters such as flame propagation speed and morphological structure synchronously, and this method lacks the ability of joint analysis of continuous parameters. The invention patent CN103792090B provides a diesel engine in-cylinder combustion visualization testing device, but the subsequent image processing does not introduce deep learning algorithm, resulting in insufficient flame boundary recognition accuracy and inability to realize time sequence tracking of flame movement.

[0003] The above-mentioned technology has the following defects: first, the traditional image processing method is limited by two-dimensional image analysis and cannot reconstruct the three-dimensional structure of the flame from a single view image, resulting in invalid spatial position prediction; second, the existing deep learning model mainly focuses on the basic identification of whether the flame exists, and does not combine dynamic analysis techniques such as optical flow method and background subtraction, so it cannot capture the high-frequency deformation and motion trajectory of the flame in the engine combustion process; finally, the feature parameter extraction link lacks a multi-modal fusion mechanism, such as isolated processing of parameters such as flame area, propagation speed and color distribution, which hinders the comprehensive evaluation of the combustion state. The reason is that the flame has instantaneous deformation characteristics and complex background interference, and the traditional algorithm is difficult to stably segment the boundary; three-dimensional reconstruction needs to rely on multi-view synchronous acquisition, which is difficult to implement in the closed cylinder of the engine; at the same time, dynamic parameter extraction needs to combine time sequence analysis and spatial modeling, and the existing technology does not establish an end-to-end processing flow. Therefore, there is an urgent need for a method that combines deep learning and multi-dimensional image processing to realize automatic recognition, three-dimensional positioning and multi-parameter joint extraction of the flame. SUMMARY

[0004] To solve the above technical problems, the present application provides a flame image recognition and feature extraction method and system to solve the problems existing in the prior art.

[0005] In the first aspect, to achieve the above-mentioned purpose, the present application provides a flame image recognition and feature extraction method and system, comprising the following steps:

[0006] The flame images obtained by the engine combustion visualization test are preprocessed, including background noise reduction and boundary enhancement;

[0007] Based on the preprocessed images, a mask of the flame region is generated by an image segmentation algorithm;

[0008] The flame movement is tracked by optical flow method and background subtraction, and the dynamic behavior of the flame is analyzed;

[0009] The spatial position of the flame is reconstructed in three dimensions, including monocular depth estimation and binocular vision parallax calculation;

[0010] A convolutional neural network model is trained based on the segmented flame images, and the continuity parameters and discrete parameters of the flame are predicted by the regression branch and the classification branch respectively;

[0011] The feature parameters of the flame are extracted, including area, propagation speed, color distribution and spatial position, to complete the flame recognition and parameter analysis.

[0012] Optionally, the image segmentation algorithm includes coarse segmentation and fine segmentation;

[0013] The coarse segmentation generates a preliminary mask of the flame region using an adaptive threshold method;

[0014] The fine segmentation takes the preliminary mask as a prior input and extracts the accurate flame boundary in combination with a U-Net semantic segmentation network.

[0015] Optionally, the tracking process of the flame movement includes:

[0016] The flame foreground region is separated by background subtraction to generate a binary mask;

[0017] The optical flow method is applied to calculate the flame motion vector for continuous frame images to analyze the propagation speed and direction.

[0018] Optionally, the process of three-dimensional reconstruction includes:

[0019] The MonoDepth2 model is used for monocular depth estimation to predict the depth information of the flame;

[0020] Based on the parallax calculation and triangulation of binocular vision, the flame three-dimensional point cloud is reconstructed and the spatial coordinates are calculated.

[0021] Optionally, the convolutional neural network model takes ResNet-50 as the backbone network;

[0022] The regression branch is used to output the continuity parameters of the flame;

[0023] The classification branch is used to output the discrete parameters of the flame.

[0024] Optionally, the feature parameter extraction process comprises:

[0025] calculating the geometric center and the circumscribed cube of the flame point cloud;

[0026] fusing the flame area, speed, color histogram and edge features, and performing final recognition through an SVM classifier.

[0027] In a second aspect, the present application further provides a flame image recognition and feature extraction system for implementing the flame image recognition and feature extraction method and system, and the system comprises:

[0028] an image preprocessing module for performing background noise reduction and boundary enhancement on the flame image obtained through engine combustion visualization testing;

[0029] an image segmentation module for generating a mask of the flame region based on the preprocessed image;

[0030] a motion tracking module for analyzing the dynamic behavior of the flame through optical flow and background subtraction;

[0031] a three-dimensional reconstruction module for determining the spatial position of the flame through monocular depth estimation and binocular vision parallax calculation;

[0032] a parameter prediction module for training a convolutional neural network model based on the segmented flame image, and predicting the continuity parameters and discrete parameters of the flame through regression branches and classification branches, respectively;

[0033] a feature analysis module for extracting the area, propagation speed, color distribution and spatial position parameters of the flame.

[0034] In a third aspect, the present application further provides a computer terminal device, comprising:

[0035] one or more processors;

[0036] a memory coupled to the processor, configured to store one or more programs;

[0037] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the flame image recognition and feature extraction method and system in the first aspect.

[0038] In a fourth aspect, the present application further provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the flame image recognition and feature extraction method and system in the first aspect.

[0039] In a fifth aspect, the present application further provides a computer program product comprising a computer program, wherein the computer program, when executed by a processor, implements the steps of the flame image recognition and feature extraction method and system of the first aspect.

[0040] Compared with the prior art, the present application has the following advantages and technical effects:

[0041] The flame image recognition and feature extraction method and system provided by the present application realize automatic recognition and three-dimensional spatial positioning of engine combustion flames, breaking through the limitation of traditional methods that cannot predict flame position from two-dimensional images; through joint regression and classification branches of a deep learning model, continuous parameters and discrete parameters of the flame are synchronously output, and multi-feature extraction such as flame propagation speed and temperature distribution is completed; dynamic flame tracking is realized by combining the optical flow method and background subtraction, which is suitable for high deformability of the combustion process; the three-dimensional reconstruction technology is used to accurately calculate the flame centroid and circumscribed cube, and the spatial analysis capability is enhanced; finally, an end-to-end feature extraction process is established, reducing manual intervention and increasing image processing quality and precision. BRIEF DESCRIPTION OF DRAWINGS

[0042] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of the present application illustrated in the drawings and their descriptions are used to explain the present application and are not intended to limit the present application. In the drawings:

[0043] Figure 1 A specific flowchart of the flame parameter prediction of the embodiment of the present application is shown in Figure 4.

[0044] Figure 2 A comprehensive flowchart of the neural network model of the flame recognition of the embodiment of the present application is shown in Figure 5. DETAILED DESCRIPTION

[0045] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0046] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a group of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in a different order.

[0047] Embodiment one

[0048] As shown in Figure 1, the present embodiment provides a flame image recognition and feature extraction method and system, which comprises: Figure 1

[0049] ​The flame image obtained by engine combustion visualization test is preprocessed, including background noise reduction and boundary enhancement;

[0050] Based on the preprocessed image, the mask of the flame area is generated by image segmentation algorithm;

[0051] The flame movement is tracked by optical flow method and background subtraction, and the dynamic behavior of the flame is analyzed;

[0052] The spatial position of the three-dimensional reconstructed flame is reconstructed, including monocular depth estimation and binocular vision parallax calculation;

[0053] Based on the segmented flame image, a convolutional neural network model is trained, and the continuity parameters and discrete parameters of the flame are predicted by the regression branch and the classification branch respectively;

[0054] The feature parameters of the flame are extracted, including area, propagation speed, color distribution and spatial position, and the flame recognition and parameter analysis are completed.

[0055] Specifically, the detailed process of the above method includes: (1) flame image preprocessing: the flame image obtained by visualization test is preprocessed, including background noise reduction, boundary enhancement, etc.

[0056] (2) Flame recognition algorithm building: a flame recognition algorithm is established, which is combined with an image segmentation algorithm to perform image segmentation on the preprocessed image, create a flame image dataset, and provide a basis for subsequent neural network establishment and deep learning;

[0057] (3) Flame automatic recognition model building: the flame movement is tracked by optical flow method and background subtraction, and the three-dimensional reconstruction technology is used to establish the in-cylinder spatial coordinate system and the image coordinate system, so as to determine the spatial position of the flame, and provide a basis for subsequent multi-working condition verification and parameter analysis;

[0058] (4) Flame parameter prediction: based on the preprocessed and recognized image,

[0059] After learning using convolutional neural network, regression algorithm is built and CNN is used to extract image features, ResNet is used as backbone, and regression value and classification result are output simultaneously through branch structure. During model building, pre-trained ResNet-50 (without last fully connected layer) is used, and feature dimension is output through Global Average Pooling on this basis; fully connected layer and activation function are added to establish regression branch and classification branch, respectively output continuous paper of coordinate and flame classification function, and loss function of corresponding branch is obtained, finally realize the separate prediction method of continuity parameters and discrete parameters of flame;

[0060] (5) Finally, the trained deep learning model is obtained, and the flame image obtained from the engine visualization test is input. The model is used to extract the relevant parameters of the flame image.

[0061] To address the challenges of flame region identification and feature parameter extraction in engine combustion visualization, improved image processing methods, particularly deep learning neural networks, are employed to predict flame spatial location and feature parameters. This significantly enhances the efficiency of flame identification, feature parameter extraction, and analysis, while also improving image processing quality and accuracy. Neural networks, especially convolutional neural networks (CNNs), can automatically extract features from complex images, such as flame shape, temperature distribution, and color changes, and integrate this information to provide a more comprehensive analysis of combustion status. Through deep learning model training, neural networks can automatically learn subtle features in images, improving analysis accuracy and providing earlier warnings when minor anomalies are detected, allowing for preventative measures to avoid potential malfunctions. Engine combustion visualization generates vast amounts of data. Compared to the time-consuming and error-prone nature of manual analysis, deep learning models can automatically classify, label, and extract features from large volumes of image data, thereby improving data processing efficiency. Finally, deep learning can learn and improve with the accumulation of more data. As experimental data is continuously updated, neural network models become increasingly accurate and adaptable to more complex or different types of combustion environments.

[0062] As one implementation method in this embodiment, the image segmentation algorithm includes coarse segmentation and fine segmentation;

[0063] The coarse segmentation uses an adaptive thresholding method to generate a preliminary mask for the flame region;

[0064] The fine segmentation uses the initial mask as prior input and combines it with the U-Net semantic segmentation network to extract accurate flame boundaries.

[0065] Specifically, flame images acquired by an engine combustion visualization device are used as the research object. First, the images are preprocessed, including dynamic background noise reduction, brightness and contrast enhancement, and grayscale conversion to simplify subsequent image processing. An image segmentation algorithm is developed, using Otsu's method or adaptive thresholding on the preprocessed flame images. This automatically finds the optimal threshold from the grayscale image, generating a preliminary binary image of the flame region. This allows for coarse localization of the entire flame region, generating a flame mask. This mask can be input into the subsequent segmentation network as attention / guiding input, or it can mask non-flame regions, reducing the probability of misidentification. Furthermore, a fine flame boundary is extracted using the U-Net semantic segmentation model. The preliminary mask used for coarse flame localization is used as a priori for the U-Net semantic segmentation network, compared with the original... Figure 1 The input method enhances the ability to accurately extract flame areas.

[0066] After the flame image is preprocessed and segmented, a neural network model is trained from the existing flame dataset to identify the flame region in the image. A convolutional neural network (CNN) is used to extract the features of the image, and the segmented image is used for training. The existing dataset is augmented (rotated, flipped, cropped, etc.) to improve the robustness of the model. Before training the neural network for flame recognition, the training data needs to be normalized, and then the activation function of the hidden layer of the CNN neural network is selected. Generally, the ReLU function is used, which is simple and efficient, and is commonly used in CNN hidden layers to avoid gradient disappearance problems, as shown in the following formula:

[0067] f(x) = max(0, x)

[0068] During the training process, in order to monitor the training (learning) process and prevent "overfitting", improve the simulation performance and generalization ability of the network model, the sample data needs to be randomly divided into three parts: training samples (60%), test samples (20%), and prediction samples (20%). Through convolutional neural network learning, regression algorithms and classification algorithms are used for continuous parameters and discrete types, respectively.

[0069] As an embodiment in this embodiment, the tracking process of the flame movement includes:

[0070] Separate the flame foreground region by background subtraction to generate a binary mask;

[0071] Apply the optical flow method to calculate the flame motion vector for continuous frame images to analyze the propagation speed and direction.

[0072] Specifically, a flame image dataset is established, and the optical flow method and background subtraction are used for flame motion tracking. First, background subtraction is used to generate a foreground mask for each frame of image, separate the flame region, eliminate the static background and the region unrelated to the flame, output the binary image of the foreground region, and the flame region is marked as white and the background region is marked as black. Then, further morphological operation is performed on the mask to remove noise and fill possible holes, ensuring the continuity of the flame region. Based on the foreground mask, the position and shape of the flame region are extracted, and the optical flow method is used for motion analysis on the target region. Input consecutive frames to calculate the motion vector of the flame region; by masking the flame region mask obtained by background subtraction, the corresponding motion vector is extracted, the motion vector of the background is ignored, and the optical flow of the flame region is calculated. The speed and motion direction of the flame are calculated and it is judged whether there is an ordered and continuous motion.

[0073] Finally, the results of background subtraction and optical flow are combined to analyze the motion information of each frame. Based on the analysis of the flame motion characteristics (such as orderly rising, flame speed, etc.), it is further determined whether it is a flame or other dynamic source (such as light, reflection, etc.). The entire process is time-sequenced, using background subtraction to process each frame of image, eliminating the background static area and marking the flame area; using optical flow to calculate the motion vector of the area, analyzing the dynamic behavior of the flame, and comprehensively analyzing the optical flow data of multiple frames to determine the orderliness of the motion (such as the upward floating of the flame), and further confirming whether it is a flame. Since background subtraction provides a rough segmentation of the foreground area, the weight is higher during cooperative processing; optical flow provides a fine analysis of the motion characteristics, and the weight is lower, but is used to determine whether there is an orderly motion.

[0074] As an embodiment in this embodiment, the process of three-dimensional reconstruction includes:

[0075] MonoDepth2 model is used for monocular depth estimation to predict the depth information of the flame;

[0076] Based on the disparity calculation and triangulation of binocular vision, the three-dimensional point cloud of the flame is reconstructed and the spatial coordinates are calculated.

[0077] Specifically, in order to realize the three-dimensional structure reconstruction of the flame, one of the core steps is to accurately estimate the depth information of the flame in the image. Considering the instability, high deformability of the flame and the difficulty of obtaining multi-view synchronous images in actual shooting process, this project adopts monocular depth estimation method which is more practical and efficient, and uses it as a depth prediction module to further reconstruct the three-dimensional spatial distribution of the flame. On this basis, MonoDepth2 model is selected as the depth estimation core network. First, MonoDepth2 is based on ResNet encoder-decoder structure, combined with self-supervised learning mechanism, which can be trained without GT depth label, suitable for scenes such as flame which are difficult to obtain real depth data; second, MonoDepth2 has excellent performance on KITTI and other datasets, has good generalization ability, and is suitable for migration to flame image scene; finally, MonoDepth2 has certain adaptability to dynamic objects (such as pedestrians, vehicles), and is suitable for processing flame images which are also dynamic and complex.

[0078] The 3D center coordinates of the flame are calculated using the projection matrix and the disparity map obtained by camera calibration. To achieve three-dimensional reconstruction and center coordinate estimation of the flame, one of the key steps is to generate an accurate disparity map. The binocular stereo vision method is used to generate the disparity map. The depth of each pixel point is recovered by using the disparity difference between the two camera images. The SGM (Semi-Global Matching) is used for disparity calculation. Based on each pixel in the left image, the matching block is searched in the corresponding scanning line of the right image, and the disparity value d is calculated. The following is the disparity calculation logic of the binocular stereo vision method:

[0079] Let the left and right images be I L (u,v), I R (u,v), the disparity d(u,v) is the horizontal offset of the pixel (u,v) in the left image and the corresponding matching point in the right image. In the corrected binocular image, the matching process is a one-dimensional search:

[0080] d(u,v) = u - u'

[0081] Where: u' is the horizontal coordinate of the matching pixel in the right image.

[0082] And the depth Z is calculated by triangulation:

[0083]

[0084] Where: f is the camera focal length (unit: pixels); B is the binocular baseline (the physical distance between the two cameras); d(u,v) is the pixel disparity value.

[0085] Finally, the depth value is back-projected to three-dimensional coordinates using the camera intrinsic matrix K:

[0086]

[0087] The mean of all three-dimensional points is calculated as the geometric center (centroid) of the flame:

[0088]

[0089] The spatial features such as the circumscribed cube and the center point of the flame are extracted. To obtain the most compact and rotationally aligned cube, we use the principal component analysis (PCA) dimensionality reduction fitting method to extract the main direction of the flame point cloud and construct the minimum circumscribed cube (OBB).

[0090] The flame point cloud is preprocessed to remove background points and noise points, and the point cloud is cleaned by voxel filtering and distance threshold screening. Then the flame centroid (center point) is calculated:

[0091]

[0092] wherein P i ∈R 3 represents a three-dimensional point in the point cloud.

[0093] After centering, PCA principal direction analysis is performed. First, covariance analysis is performed on the centered point cloud:

[0094]

[0095] Then, the eigenvalues and eigenvectors are solved to obtain the principal direction axis (principal component). The first principal component (vector corresponding to the largest eigenvalue) is the maximum expansion direction of the flame; the remaining principal components constitute an orthogonal basis, which is used to construct a local coordinate system.

[0096] Project the point cloud and calculate the boundary. Project the point cloud onto the principal direction coordinate system (PCA basis); and calculate the minimum and maximum projection values on each axis to construct the minimum enclosing cube boundary.

[0097] Finally, the OBB is constructed and the world coordinate system is restored. According to the principal axis direction, the boundary range and the center point, the OBB is generated; using the rotation matrix, the OBB is transformed from the principal direction coordinate system back to the original world coordinate system.

[0098] As an implementation manner in the embodiment, the convolutional neural network model takes ResNet-50 as a backbone network;

[0099] The regression branch is used to output the continuity parameter of the flame;

[0100] The classification branch is used to output the discrete parameter of the flame.

[0101] As an implementation manner in the embodiment, the feature parameter extraction process includes:

[0102] Calculate the geometric center and the circumscribed cube of the flame point cloud;

[0103] Fuse the flame area, speed, color histogram and edge features, and perform final recognition through an SVM classifier.

[0104] Specifically, flame feature parameters such as flame area, flame boundary side length, average / maximum speed of flame propagation, color distribution, and vibration frequency are extracted. Based on the flame feature parameters, different features (such as area, speed, color histogram, and edge shape) are combined through multi-feature fusion, and a machine learning classification algorithm is used to accurately extract the flame boundary for final flame recognition. In the multi-feature fusion process, feature-level fusion is used to combine different features into a feature vector, which is then input into the classifier for training and prediction; decision-level fusion is used to use multiple classifiers to make decisions based on single features, and finally the final result is obtained through weighted voting and decision fusion.

[0105] Support Vector Machine (SVM) is used as the classifier and multiple features are used for training. The following is the feature fusion logic based on the SVM classifier:

[0106] First, feature extraction and normalization are performed to construct a feature vector. The above four main features (area, speed, color histogram, and shape) are extracted; and each feature is normalized or standardized to ensure that the scales of different features are consistent and to avoid the excessive influence of certain features on the classification result. The normalized features are spliced into a complete feature vector X:

[0107] X = [Flame Area, Flame Velocity, Color Histogram, Edge Features]

[0108] This feature vector will be used as the input of the SVM classifier.

[0109] On this basis, the labeled flame and non-flame samples are used to train the SVM model, so that the SVM can learn the optimal decision boundary between flame and non-flame under different feature combinations. The feature vector of the new flame image or video frame is input into the trained SVM classifier to obtain the recognition result.

[0110] Through the recognition of flame shape and the extraction of feature parameters, the combustion mechanism is further analyzed, the depth of learning of combustion excitation is optimized, and the recognition and analysis accuracy is optimized. By combining the spectral information of the flame of the hyperspectral camera with the traditional visual image data, more dimensional features can be provided for the recognition and analysis of the flame, so that the characteristics of the flame can be more comprehensively understood. Thus, a correlation model between spectral characteristics and combustion process is established, and the depth of learning of combustion visualization and flame recognition is optimized.

[0111] The specific process of flame parameter prediction is as follows: Figure 1The neural network model comprehensive flowchart of the flame recognition is shown in Figure 2 .

[0112] (1). The specific flowchart of the flame parameter prediction is shown in Figure 1 . The flame image obtained by the engine combustion visualization device is taken as the research object, and the image is preprocessed first, and the dynamic background noise reduction and boundary enhancement are performed on multiple flame images.

[0113] (2). After the flame image preprocessing and segmentation, the flame image dataset is established, the flame motion tracking is performed by using the optical flow method and the background subtraction, the flame three-dimensional structure reconstruction algorithm is written, the flame spatial position is determined, the flame feature parameters are extracted, the flame recognition method is established, the flame boundary is accurately extracted, and the flame features are analyzed.

[0114] (3). The artificial neural network prediction model is established, including output / input parameters, normalization processing, network structure selection, network training and the like. The dataset composed of the flame segmentation results is used for the training of the machine learning. In order to monitor the training (learning) process and prevent the overfitting, improve the simulation performance and the generalization ability of the network model, the sample data needs to be randomly divided into three parts, i.e., the training sample (60%), the test sample (20%) and the prediction sample (20%). Through the convolutional neural network learning, the regression algorithm and the classification algorithm are used to predict the continuous parameters and the discrete parameters respectively. The predicted data is compared with the actual analysis or simulation data, and the precision verification is performed.

[0115] (4). Through the recognition of the flame shape and the extraction of the feature parameters, the combustion mechanism is further analyzed, the combustion excitation is deeply learned, the recognition analysis precision is optimized, the correlation model between the spectral characteristics and the combustion process is established, and the deep learning of the combustion visualization and the flame recognition are optimized.

[0116] (5). According to the above flowchart, the high-load point emission data of the real ship engine can be predicted.

[0117] Compared with the prior art, the embodiments of the present application have the beneficial effects including:

[0118] (1) The traditional combustion visualization processing image in the invention patent "A diesel engine in-cylinder spray and combustion visualization testing device and testing method" (application number CN201310293341.8, application publication date January 27, 2016, authorized publication number CN103792090B) cannot determine the flame space position and track the flame movement through image processing alone, and cannot identify the specific parameters of the flame from two-dimensional images. Compared with the above-mentioned file, the present invention introduces deep learning into flame identification and feature extraction, which can take advantage of deep learning for image processing, not only improving the efficiency and accuracy of analysis, but also helping to find potential problems and optimize the combustion process. With the increase of data volume, the deep learning model can continuously evolve and improve, providing strong support for engine performance optimization and fault warning;

[0119] (2) Compared with the invention patent "A flame identification method and a readable storage medium" (Chinese patent number CN202411005602.6, application publication date December 6, 2024, application publication number CN119091368A) and the invention patent "A flame identification method based on deep learning" (application number CN202410869983.6, application publication date October 1, 2024, authorized publication number CN118736491A) for identifying whether the flame exists, the deep learning of the present invention can better deal with the problems in the engine combustion visualization process, and can more effectively identify the flame area, determine the accurate position of the flame, extract the flame feature parameters, and improve the efficiency of the engine combustion visualization test;

[0120] (3) The present invention has wide application prospects and can be used for internal combustion engine combustion visualization tests in the fields of automobiles, ships, aerospace, etc. At the same time, the present invention also has certain reference significance for systems that need to identify combustion flames and extract feature parameters in other fields.

[0121] Embodiment Two

[0122] In this embodiment, a computer terminal device is provided, comprising:

[0123] one or more processors;

[0124] a memory coupled to the processor, for storing one or more programs;

[0125] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the above-mentioned flame image identification and feature extraction method and system.

[0126] In the embodiment, a computer readable storage medium is also provided, and the computer readable storage medium stores a computer program. The computer program is executed by a processor to implement the steps of the flame image recognition and feature extraction method and system.

[0127] In the embodiment, an electronic device is also provided, and the electronic device comprises a memory and a processor. The memory stores a computer program, and the processor is configured to execute the computer program to perform the steps of the flame image recognition and feature extraction method and system.

[0128] In the embodiment, a computer program product is also provided, and the computer program product comprises a computer program. The computer program is executed by a processor to implement the steps of the flame image recognition and feature extraction method and system.

[0129] The above program can be executed in a processor, or can also be stored in a memory (or a computer readable medium). The computer readable medium includes permanent and non-permanent, removable and non-removable media, and can be implemented by any method or technology to store information. The information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technologies, compact disc read only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device.

[0130] The computer program can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are performed on the computer or other programmable device to generate a computer implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more flows Figure 1 The functions specified in one or more flows or one or more blocks and / or blocks Figure 1 The functions specified in one or more flows or one or more blocks and / or blocks

[0131] In the embodiment, such a device or system is provided. The system is called a flame image recognition and feature extraction system, which comprises:

[0132] The image preprocessing module is used for background noise reduction and boundary enhancement of the flame image obtained by engine combustion visualization test.

[0133] an image segmentation module configured to generate a mask of the flame region based on the preprocessed image;

[0134] a motion tracking module configured to analyze dynamic behavior of the flame by optical flow and background subtraction;

[0135] a three-dimensional reconstruction module configured to determine spatial position of the flame by monocular depth estimation and binocular vision parallax calculation;

[0136] a parameter prediction module configured to train a convolutional neural network model based on the segmented flame image, and to predict continuous parameters and discrete parameters of the flame by regression branch and classification branch respectively;

[0137] a feature analysis module configured to extract area, propagation speed, color distribution and spatial position parameters of the flame.

[0138] As an embodiment of the present embodiment, the image segmentation module comprises:

[0139] a coarse segmentation unit configured to generate a preliminary mask of the flame region by using an adaptive threshold method;

[0140] a fine segmentation unit configured to extract an accurate flame boundary by using the preliminary mask as a prior input and combining a U-Net semantic segmentation network.

[0141] As an embodiment of the present embodiment, the motion tracking module comprises:

[0142] a foreground separation unit configured to generate a binary mask of the flame region by background subtraction;

[0143] an optical flow analysis unit configured to calculate flame motion vectors and analyze propagation characteristics for consecutive frame images.

[0144] As an embodiment of the present embodiment, the three-dimensional reconstruction module comprises:

[0145] a monocular depth estimation unit configured to predict flame depth information by using a MonoDepth2 model;

[0146] a binocular reconstruction unit configured to reconstruct three-dimensional point cloud and spatial coordinates of the flame based on parallax calculation and triangulation.

[0147] As an embodiment of the present embodiment, the parameter prediction module comprises:

[0148] a ResNet backbone unit configured to build a convolutional neural network with ResNet-50 as the core;

[0149] a regression output unit configured to generate a prediction result of the continuous parameters of the flame;

[0150] A classification output unit is configured to generate a prediction result of the flame discrete parameter.

[0151] As an embodiment in the present embodiment, the feature analysis module comprises:

[0152] A geometry calculation unit is configured to calculate a geometric center and an outer-enclosing cube of the flame point cloud.

[0153] A multi-feature fusion unit is configured to integrate the flame area, speed, color histogram and edge feature, and complete the final recognition through an SVM classifier.

[0154] The system or device is used to realize the functions of the method in the above-mentioned embodiments. Each module in the system or device corresponds to each step in the method, and has been described in the method and will not be described here.

[0155] Through the above-mentioned embodiments, the problems of flame image recognition and feature extraction in the related art are solved, thereby being able to guarantee to solve the problems in the prior art.

[0156] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A flame image recognition and feature extraction method and system, characterized in that, The method comprises the following steps: The flame image obtained by engine combustion visualization test is preprocessed, including background noise reduction and boundary enhancement; Based on the preprocessed image, the mask of the flame region is generated by image segmentation algorithm; The flame movement is tracked by optical flow method and background subtraction to analyze the dynamic behavior of the flame; The spatial position of the flame is reconstructed in three dimensions, including monocular depth estimation and binocular vision parallax calculation; A convolutional neural network model is trained based on the segmented flame image, and the continuity parameters and discrete parameters of the flame are predicted by regression branch and classification branch respectively; The feature parameters of the flame are extracted, including area, propagation speed, color distribution and spatial position, to complete flame recognition and parameter analysis.

2. The method of claim 1, wherein, The image segmentation algorithm comprises coarse segmentation and fine segmentation; The coarse segmentation generates a preliminary mask of the flame region by using an adaptive threshold method; The fine segmentation takes the preliminary mask as a prior input and extracts the accurate flame boundary by combining a U-Net semantic segmentation network.

3. The method of claim 1, wherein, The tracking process of the flame movement comprises: The flame foreground region is separated by background subtraction to generate a binary mask; The optical flow method is applied to calculate the flame motion vector for continuous frame images to analyze the propagation speed and direction.

4. The method of claim 1, wherein, The process of three-dimensional reconstruction comprises: MonoDepth2 model is used for monocular depth estimation to predict the depth information of the flame; Based on binocular vision parallax calculation and triangulation, the flame three-dimensional point cloud is reconstructed and the spatial coordinates are calculated.

5. The method of claim 1, wherein, The convolutional neural network model takes ResNet-50 as the backbone network; The regression branch is used to output the continuity parameters of the flame; The classification branch is used to output the discrete parameters of the flame.

6. The method of claim 1, wherein, The extraction process of the feature parameters comprises: The geometric center and circumscribed cube of the flame point cloud are calculated; The flame area, speed, color histogram and edge features are fused, and the final identification is performed by an SVM classifier.

7. A flame image recognition and feature extraction system characterized by, The system comprises: An image preprocessing module for background noise reduction and boundary enhancement of the flame image obtained by engine combustion visualization test; An image segmentation module for generating a mask of the flame region based on the preprocessed image; A motion tracking module for analyzing the dynamic behavior of the flame by optical flow method and background subtraction; A three-dimensional reconstruction module for determining the spatial position of the flame by monocular depth estimation and binocular vision parallax calculation; A parameter prediction module for training a convolutional neural network model based on the segmented flame image, and predicting the continuity parameters and discrete parameters of the flame by regression branch and classification branch respectively; A feature analysis module for extracting the area, propagation speed, color distribution and spatial position parameters of the flame.

8. A computer terminal device, characterized by One or more processors; Memory coupled to the processor for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the method of any one of claims 1-6. The computer program is executed by the processor to implement the steps of the method of any one of claims 1-6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1-6.

10. A computer program product comprising a computer program, characterized in that, ​

Images