An optical smoke detection method combining prior knowledge and feature classification

Abstract

The present application relates to a kind of prior knowledge and feature classification combined optical smoke detection method, comprising: based on the gray feature of smoke from target image in the candidate region of smoke segmentation;The candidate region of smoke is carried out morphological processing, in combination with the shape information of smoke and the height information of smoke, and the first candidate region of smoke is screened out;Based on the gray feature, texture feature and gradient feature of image, the multi-dimensional feature descriptor of candidate region of smoke is extracted, is sent into the trained classifier and is judged, and the second candidate region of smoke is screened out;The intersection of first candidate region of smoke and second candidate region of smoke is selected as the confirmed candidate region of smoke;Edge extraction is carried out to the confirmed candidate region of smoke, fusion, obtains the position information of smoke region, to obtain the final smoke detection result output.This algorithm has low complexity, occupies less hardware resources, can quickly realize smoke detection, and is suitable for the operation platform of limited hardware resources.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to an optical smoke detection method that combines prior knowledge and feature classification. Background Technology

[0002] Currently, smoke detection and recognition technologies are mainly focused on fire detection. The purpose of smoke detection is to identify fires earlier and gain valuable time for firefighting. Furthermore, the applications are mostly in video surveillance scenarios with fixed backgrounds, and research primarily focuses on video smoke detection and recognition technologies that utilize the color, texture, turbulence, and drift characteristics of smoke.

[0003] However, there is a lack of effective detection methods for smoke interference near the water surface against a sea and sky background. Furthermore, conventional video smoke detection and recognition techniques based on characteristics such as color, texture, turbulence, and drift are unsuitable for optical smoke detection against sea and sky backgrounds due to differences in image features (e.g., the acquired images are only grayscale and cannot utilize color information). Additionally, existing smoke recognition algorithms suffer from high computational complexity and long processing times for feature extraction and classification, requiring significant storage space. Platforms with limited hardware resources cannot meet the operational requirements of existing smoke recognition algorithms. Summary of the Invention

[0004] Based on the above analysis, the present invention aims to disclose an optical smoke detection method that combines prior knowledge and feature classification to solve the technical problem of optical smoke detection against a sea and sky background.

[0005] This invention discloses an optical smoke detection method that combines prior knowledge and feature classification, comprising:

[0006] For a target image containing optical smoke interference against a sea and sky background, smoke candidate regions are segmented from the target image based on smoke grayscale features;

[0007] Morphological processing was performed on the candidate smoke regions, and the first type of candidate smoke regions was selected by combining smoke shape information and smoke height information.

[0008] Based on the grayscale features, texture features and gradient features of the image, multidimensional feature descriptors are extracted from the smoke candidate regions, which are then fed into a trained classifier for judgment, and the second type of smoke candidate regions are selected.

[0009] The intersection of the first type of smoke candidate region and the second type of smoke candidate region is selected as the confirmed smoke candidate region;

[0010] Edge extraction and fusion are performed on the confirmed smoke candidate regions to obtain the location information of the smoke regions, thereby obtaining the final smoke detection result output.

[0011] Furthermore, morphological processing is performed on the segmented smoke candidate regions, and target dilation is performed in the horizontal direction to close the segmented smoke candidate regions.

[0012] Based on prior knowledge, including the shape and height of the smoke, candidate regions that do not meet the smoke characteristics are eliminated from the closed smoke candidate regions; the eliminated smoke candidate regions are then labeled as connected components to obtain the first type of smoke candidate regions.

[0013] Furthermore, when screening candidate regions for the second type of smoke,

[0014] The multidimensional feature descriptor contains 8 features: gray-level mean, gray-level variance, maximum gray-level value, gray-level difference mean, contrast, entropy, vertical gradient mean, and vertical gradient variance.

[0015] Furthermore, the classifier employs a tree-based classification algorithm.

[0016] Furthermore, the segmentation of the smoke candidate region includes:

[0017] 1) Based on the histogram information of the image, obtain the grayscale estimates of the sea surface and sky, and obtain the initial seed points for the sea surface and sky;

[0018] 2) Based on the statistical empirical value of the white highlight grayscale of the smoke region and the maximum grayscale value of the image, the initial seed points of the smoke are obtained;

[0019] 3) Determine the initial seed points for the sea surface, sky, and smoke. Cluster the image grayscale based on the clustering algorithm, segment and extract the potential regions of the image belonging to the smoke target, and obtain the smoke candidate regions.

[0020] Furthermore, the histogram is smoothed and peak points are detected to obtain grayscale estimates of the sea surface and sky, and to obtain initial seed points for the sea surface and sky.

[0021] The initial seed points for the sky and sea are the first and second extreme points of the histogram of the smoothed image;

[0022] The initial seed point for the smoke is Among them, V s is the statistical empirical value of the smoke grayscale, and max(img) is the maximum grayscale value of the image.

[0023] Furthermore, the Kmeans gray-level clustering algorithm is used to cluster the gray levels of the image to obtain the potential regions of the image belonging to the smoke target. After segmentation and extraction, the smoke candidate regions are obtained.

[0024] Furthermore, the methods for edge extraction and fusion of the confirmed smoke candidate regions include:

[0025] 1) An edge detection algorithm is used to extract the edge map of the smoke;

[0026] 2) Based on the edge map information, the smoke is confirmed and the segmentation results are supplemented to obtain the segmentation map after edge fusion;

[0027] 3) Perform morphological dilation on the segmented map after edge fusion to obtain the final edge feature fusion map;

[0028] 4) Connected component labeling is performed on the edge feature fusion map to obtain the final smoke detection result output.

[0029] Furthermore, for each confirmed smoke candidate region, the number of edge points at the corresponding position in the edge map is counted to determine its matching degree with the smoke candidate region. When the overlap rate condition is met, it is considered that the two have an intersection, and the current smoke candidate region is determined to be the smoke target region. The clustered smoke candidate region and the edge map corresponding to its position are combined to obtain the segmented map after fusion edge.

[0030] Furthermore, the segmentation map after edge fusion:

[0031]

[0032] In the formula, Seg(i,j) is the clustering segmentation map, Edge(i,j) is the edge extraction map, and Match(m) is the matching degree between the edge map of the m-th smoke candidate region and the current m-th smoke candidate region.

[0033] Overlap(m) is the overlap rate between the edge map of the m-th smoke candidate region and the current m-th smoke candidate region; m = 1, ..., boxnum; boxnum is the total number of smoke candidate regions in the clustering segmentation map; i, j are the vertical and horizontal coordinates of the image.

[0034] Specifically, the formula for calculating the matching degree is:

[0035]

[0036] The formula for calculating the overlap rate;

[0037] Overlap(m)=box[m].w*box[m].h*β

[0038] Where box[m] represents the coordinate information of the m-th smoke candidate region in the cluster segmentation graph represented by a rectangle, where box[m].bottom, box[m].top, box[m].left, and box[m].right represent the coordinates of the upper, lower, left, and right boundaries of the detection box, respectively, and box[m].w and box[m].h represent the width and height of the m-th rectangle; β is a settable scaling factor.

[0039] This invention can achieve at least one of the following beneficial effects:

[0040] This invention presents an optical smoke detection method that combines prior knowledge and feature classification. It employs a grayscale clustering algorithm to segment candidate smoke regions and uses prior knowledge, including smoke shape and height, along with feature classification based on eight multi-dimensional feature descriptors to filter candidate smoke regions, thus improving the accuracy of smoke detection. Furthermore, it combines edge features to determine and fuse the confirmed candidate smoke regions, resulting in a more accurate smoke target range.

[0041] This invention enables smoke detection without using motion features, solving the technical problem that the continuous movement of the aircraft platform causes the image background information to change constantly, making it impossible to use motion features to complete smoke detection.

[0042] The algorithm used in this invention is simple and effective, consumes few hardware resources, runs fast, and is suitable for hardware platforms with limited resources. Attached Figure Description

[0043] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0044] Figure 1 This is a flowchart of an optical smoke detection method combining prior knowledge and feature classification in an embodiment of the present invention.

[0045] Figure 2 This is a flowchart of the smoke candidate region segmentation method in an embodiment of the present invention;

[0046] Figure 3 A flowchart illustrating the edge extraction and fusion method for the smoke candidate region in this embodiment of the invention;

[0047] Figure 4 This is a flowchart illustrating the training process of a feature-based classifier in an embodiment of the present invention. Detailed Implementation

[0048] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and, together with the embodiments of the present invention, serve to illustrate the principles of the present invention.

[0049] One embodiment of the present invention discloses an optical smoke detection method that combines prior knowledge and feature classification, such as... Figure 1 As shown, it includes:

[0050] Step S101: For a target image containing optical smoke interference against a sea and sky background, segment the smoke candidate region from the target image based on the smoke grayscale features;

[0051] Step S102: Perform morphological processing on the smoke candidate regions, and filter out the first type of smoke candidate regions by combining smoke shape information and smoke height information;

[0052] Step S103: Based on the grayscale features, texture features and gradient features of the image, extract multidimensional feature descriptors for the smoke candidate regions, and send them to the trained classifier for judgment to select the second type of smoke candidate regions.

[0053] Step S104: Select the intersection of the first type of smoke candidate region and the second type of smoke candidate region as the confirmed smoke candidate region;

[0054] Step S105: Extract and fuse the confirmed smoke candidate regions to obtain the location information of the smoke regions, thereby obtaining the final smoke detection result output.

[0055] In the application scenario of this embodiment, the image content against the backdrop of sea and sky is relatively simple, mainly consisting of three regions: the sea surface, the sky, and the target. These three regions have significant differences in grayscale. To perform smoke segmentation more accurately, a clustering segmentation algorithm based on the three categories of sea surface, sky, and smoke can be used.

[0056] To cluster images more quickly and accurately, it is necessary to obtain accurate initial segmentation points (i.e., seed points) for three target categories: sea surface, sky, and smoke. Therefore, it is necessary to determine the approximate grayscale range of the sea surface, sky, and smoke. For images with a sea and sky background, since the background is simple, the histogram exhibits a bimodal characteristic. Initial segmentation points for clustering the sea surface and sky can be obtained by detecting the peak points of the histogram.

[0057] In step S101, a gray-scale clustering algorithm is used to segment the smoke candidate region based on the gray-scale brightness features of the smoke.

[0058] like Figure 2 As shown, the specific segmentation of the smoke candidate region includes:

[0059] Step S201: Based on the histogram information of the image, obtain the grayscale estimates of the sea surface and sky, and obtain the initial seed points of the sea surface and sky;

[0060] Preferably, grayscale estimates of the sea surface and sky are obtained by smoothing the histogram of the image and detecting peak points; the initial seed points of the sky and sea surface are the first and second extreme points of the histogram of the smoothed image.

[0061] When calculating the histogram of an image, since the histogram contains spikes (i.e., local peaks), it is necessary to smooth the histogram to eliminate the spikes. Histogram smoothing methods can include Gaussian smoothing, mean smoothing, etc.

[0062] Among them, the Gaussian smoothing calculation formula for histograms Where h i This represents the calculated image histogram, where r is the smoothing scale and σ is the standard deviation.

[0063] For the smoothed histogram, calculate the first two extreme points (i.e., find the first two maximum peaks). The formula for calculating the extreme points is H. max =max(H i-k ,…,H i ,…,H i+k ), i = 0, ..., 255; where H i This represents the smoothed histogram, where k is the scale interval for finding the maximum value.

[0064] Two maxima can be obtained according to the extreme point calculation formula. Combined with the prior knowledge that the gray level of the sea surface is lower and the gray level of the sky is higher, the corresponding gray levels are the gray level estimates of the sea surface and the sky. The first extreme point (i.e. the largest extreme point) is the initial seed point of the sky, val_sky, and the second extreme point (i.e. the second largest extreme point) is the initial seed point of the sea surface, val_sea.

[0065] Step S202: Based on the statistical empirical value of the white highlight grayscale of the smoke region and the maximum grayscale value of the image, obtain the initial seed points of the smoke.

[0066] If smoke is present in the image, based on the imaging characteristics of smoke, it will definitely have white highlighted areas. Statistical analysis of the grayscale range of the white highlighted areas in the smoke region shows that its grayscale value is basically in the range of V. s The above (V) s (Statistical values) By calculating the maximum gray level of the image and comparing the gray level corresponding to the maximum value with the statistical value of the smoke gray level, the seed points for smoke are identified. Combining the gray level estimates of the sea surface and sky obtained in step S201, the initial seed points for the three categories of sea surface, sky, and smoke are identified as follows:

[0067]

[0068] Among them, V s is the statistical empirical value of the smoke grayscale, and max(img) is the maximum grayscale value of the image.

[0069] Statistical empirical value V of smoke ash s A smoke image library is constructed by collecting all the smoke images to be detected, and the average grayscale value V of the smoke region in the image library is calculated. s .

[0070] Step S203: Based on the initial seed points of the sea surface, sky, and smoke, a clustering algorithm is used to cluster the image grayscale, and the potential image regions belonging to the smoke target are segmented and extracted to obtain smoke candidate regions.

[0071] Specifically, the K-means gray-level clustering algorithm is used to cluster the gray levels of the image to obtain the potential regions of the image belonging to the smoke target. After segmentation and extraction, the smoke candidate regions are obtained.

[0072] In this embodiment, since the initial seed point is obtained through calculation and is relatively accurate, accurate clustering results can be obtained with fewer iterations. Based on the clustering results, the potential image regions belonging to the smoke category are segmented and extracted to obtain smoke candidate regions.

[0073] Specifically, the screening of the first type of smoke candidate region in step S102 includes:

[0074] Morphological processing is performed on the segmented smoke candidate regions to close the segmented smoke candidate regions;

[0075] Preferably, when performing morphological processing on the segmented smoke candidate regions, in order not to affect the height information of the target, target dilation is only performed in the horizontal direction to achieve closure of the segmented candidate regions.

[0076] After morphological processing, candidate smoke regions are processed using prior knowledge, including the shape and height of the smoke. Closed candidate smoke regions that do not meet the characteristics of smoke are eliminated. For example, regions with excessively high or low heights, too few pixels, or too many straight edges not present in the smoke are removed. Connectivity labeling is then applied to the eliminated candidate smoke regions to obtain the first type of smoke candidate regions.

[0077] Specifically, in step S103, during the selection of the second type of smoke candidate regions, the smoke target processed in this embodiment exhibits the following color characteristics in the optical image: the presence of bright gray areas, gradual gray-level changes, and a generally high gray-level value. Its shape is also varied, ranging from approximately elliptical to approximately cloud-like. Furthermore, the texture of smoke is relatively simpler than other targets such as ships, exhibiting strong directionality and weak horizontal edges, whereas ship hulls have strong horizontal edges. Based on the above analysis of smoke characteristics, gray-level features, texture features, and gradient features of the smoke image can be extracted as final features for classification. These features are then used by a trained classifier to select the second type of smoke candidate regions.

[0078] Specifically, the multidimensional feature descriptor based on grayscale feature extraction is the three-dimensional feature of the image's grayscale mean, variance, and maximum value;

[0079] Based on the characteristics of smoke images, which have high grayscale brightness and grayscale gradient, while ship-type targets have low overall grayscale, this embodiment mainly extracts the three-dimensional features of the image, namely the mean, variance, and maximum value, in terms of grayscale features.

[0080] Here, the mean refers to the average value of the image's gray levels, the variance refers to the variance of the image's gray levels, and the maximum value refers to the maximum value of the image's gray levels. The definitions are as follows:

[0081] Mean:

[0082]

[0083] variance:

[0084]

[0085] Maximum value:

[0086]

[0087] Where M and N are the image width and height, and P ij This represents the grayscale value of a pixel.

[0088] Specifically, the multidimensional feature descriptor extracted based on texture features is the three-dimensional features of the image: gray-level difference mean, contrast, and entropy.

[0089] Texture features describe the surface properties of an image or the object corresponding to an image region, and quantify the features of gray level changes within the region.

[0090] The texture based on smoke is relatively simple compared to ship-type targets. In this embodiment, the texture features are extracted using the gray-level difference method.

[0091] Specifically, let there be a pixel (m, n) in the image, and the values ​​of this pixel and its neighboring pixels... The grayscale difference is:

[0092]

[0093] in, This is called gray-level difference. Suppose there are m possible levels of gray-level difference values, and we need to find... A histogram. The histogram can be used to calculate... The probability of taking a value is p(k), where k is the grayscale difference. A larger p(k) indicates a coarser texture, and a smaller p(k) indicates a finer texture.

[0094] This embodiment primarily extracts three texture features: average gray-level difference, contrast, and entropy. These are defined as follows:

[0095] Mean of grayscale difference:

[0096]

[0097] Contrast:

[0098]

[0099] entropy:

[0100]

[0101] The multidimensional feature descriptor based on gradient feature extraction is a two-dimensional gradient feature of the image, consisting of the mean and variance of the vertical gradient.

[0102] Given that smoke images exhibit strong directionality and weak horizontal edges, while ship-like targets possess strong horizontal edges, this embodiment employs the vertical gradient features of the image. The mean and variance of the vertical gradient of the image are calculated as the extracted 2D gradient features.

[0103] The Sobel operator is used to obtain the template for the vertical gradient, as shown in the formula.

[0104]

[0105] Obtain the vertical gradient G y The mean and variance of the vertical gradient of the image are calculated using the following formulas:

[0106] Vertical gradient mean:

[0107]

[0108] In the formula, G y (i,j) represents the vertical gradient of the coordinate point (i,j); N is the image width;

[0109] Vertical gradient variance:

[0110]

[0111] The two-dimensional features of the vertical gradient mean and variance are used as gradient features.

[0112] The final extracted multidimensional feature descriptor contains 8 features: gray-level mean, gray-level variance, maximum gray-level value, gray-level difference mean, contrast, entropy, vertical gradient mean, and vertical gradient variance.

[0113] For each suspected smoke target area block, the eight multi-dimensional feature descriptors are extracted and fed into the trained classifier to obtain the classifier's discrimination result. The discrimination method is to output 1 if it is a target and 0 if it is not a target. The target area identified as smoke is taken as the second type of smoke candidate area.

[0114] In step S104, the intersection of the first type of smoke candidate region and the second type of smoke candidate region is selected as the confirmed smoke candidate region.

[0115] In this step, by combining prior knowledge, including the shape and height of the smoke, with feature classification using an 8-dimensional feature descriptor, the accuracy of smoke candidate regions is further improved, thus increasing the accuracy of smoke detection.

[0116] In this embodiment, the smoke segmentation method based on grayscale clustering does not consider information such as target boundaries and contrast. Therefore, false alarms may occur when the camera is overexposed and the sky background is too bright. Furthermore, due to the varying shapes of smoke and the uneven grayscale distribution within the smoke region, exhibiting a phenomenon where some areas are bright and others are low grayscale, simply using a clustering segmentation algorithm will result in the same target being segmented into multiple sub-blocks. For these reasons, step S105 of this embodiment confirms the clustering segmentation results and further supplements and connects the segmentation results.

[0117] Specifically, such as Figure 3 As shown, the method for edge extraction and fusion of the confirmed smoke candidate regions in step S105 includes:

[0118] Step S301: Extract the edge map of the smoke using an edge detection algorithm;

[0119] Since optical smoke, which causes interference, typically only appears near the sky and horizon, the contrast between the smoke and the background is high, and edge information is present. Therefore, edge detection algorithms (such as the Canny edge extraction method) are used to extract edges from the image to obtain the edge map of the smoke.

[0120] Specifically, when extracting edges from an image, a hough transform is used to extract the sea-line area. After removing the extracted sea-line area, the remaining edges are used as the edge map of the smoke.

[0121] Step S302: Based on the edge map information, confirm the smoke and supplement the segmentation results to obtain the segmentation map after edge fusion;

[0122] For each smoke candidate region, the number of edge points at the corresponding position in the edge map is counted to determine its matching degree with the smoke candidate region. When the overlap rate condition is met, the two are considered to have an intersection, and the current smoke candidate region is determined to be the smoke target region. The clustered smoke candidate region and the edge map corresponding to its position are combined to obtain the segmented map after fusion edge.

[0123] Specifically, the segmented image after edge fusion:

[0124]

[0125] In the formula, Seg(i,j) is the clustering segmentation map, Edge(i,j) is the edge extraction map, and Match(m) is the matching degree between the edge map of the m-th smoke candidate region and the current m-th smoke candidate region.

[0126] Overlay(m) is the overlap rate between the edge map corresponding to the m-th smoke candidate region and the current m-th smoke candidate region; m = 1, ..., boxnum; boxnum is the total number of smoke candidate regions in the clustering segmentation map; i, j are the vertical and horizontal coordinates of the image.

[0127] Specifically, the formula for calculating the matching degree is:

[0128]

[0129] The formula for calculating the overlap rate;

[0130] Overlay(m)=box[m].w*box[m].h*β

[0131] Where bos[m] represents the coordinate information of the m-th smoke candidate region in the cluster segmentation graph represented by a rectangular box, where box[m].bottom, box[m].top, box[m].left, and box[m].right represent the coordinates of the upper, lower, left, and right boundaries of the detection box, respectively, and box[m].w and box[n].h represent the width and height of the m-th rectangular box; β is a settable scaling factor.

[0132] Step S303: Perform morphological dilation on the segmentation map after edge fusion to obtain the final edge feature fusion map;

[0133] Step S304: Mark the connected components of the edge feature fusion map to obtain the final smoke detection result output.

[0134] Preferably, this embodiment also discloses the training process of a feature-based classifier; such as Figure 4 include

[0135] Step S401: Establish a sample library of smoke targets, other targets, and background targets for classifier learning;

[0136] For all training images containing smoke, smoke image patches are extracted to construct a smoke image library as a positive sample library, while other targets (such as ships) and background targets are extracted as a negative sample library.

[0137] Specifically, the smoke targets processed in this embodiment exhibit the following color characteristics in optical images: the presence of bright gray areas, gradual gray-level changes, and a generally high gray-level value. Their shapes are varied, ranging from approximately elliptical to cloud-like. Furthermore, the texture of smoke is relatively simpler than other targets such as ships, exhibiting strong directionality and weak horizontal edges, whereas ship hulls have strong horizontal edges. Based on the above analysis of smoke characteristics, gray-level features, texture features, and gradient features of the smoke image can be extracted, thus serving as the final features for classification.

[0138] Step S402: Based on the grayscale features, texture features, and gradient features of the image, extract multidimensional feature descriptors for all samples in the sample library;

[0139] The multidimensional feature descriptor extracted for each sample in the sample library contains 8 features, namely gray mean, gray variance, maximum gray value, gray difference mean, contrast, entropy, vertical gradient mean, and vertical gradient variance.

[0140] Step S403: Train the classifier using multidimensional feature descriptors extracted from all samples, so that the classifier can identify smoke targets;

[0141] Classification algorithms can employ tree-based classification methods, such as the CART binary tree classification algorithm. Tree-based classification algorithms achieve fast classification due to their low complexity, low hardware resource consumption, and fast execution speed, making them suitable for resource-constrained hardware platforms. Once training is complete, a smoke target classifier can be obtained.

[0142] In summary, the optical smoke detection method combining prior knowledge and feature classification disclosed in this invention employs a grayscale clustering algorithm to segment smoke candidate regions. By combining prior knowledge, including the shape and height of the smoke, with feature classification using an 8-dimensional feature descriptor, the accuracy of smoke detection is improved. Furthermore, the confirmed smoke candidate regions are judged and fused by combining edge features to obtain a more accurate smoke target range.

[0143] The embodiments of the present invention can achieve smoke detection without using motion features, solving the technical problem that the continuous movement of the aircraft platform causes the image background information to change continuously, making it impossible to use motion features to complete smoke detection.

[0144] The algorithm used in this embodiment of the invention is simple and effective, consumes little hardware resources, runs fast, and is suitable for hardware platforms with limited resources.

[0145] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An optical smoke detection method combining prior knowledge and feature classification, characterized in that, include: For a target image containing optical smoke interference against a sea and sky background, smoke candidate regions are segmented from the target image based on smoke grayscale features; Morphological processing was performed on the candidate smoke regions, and the first type of candidate smoke regions was selected by combining smoke shape information and smoke height information. Based on the grayscale features, texture features and gradient features of the image, multidimensional feature descriptors are extracted from the smoke candidate regions, which are then fed into a trained classifier for judgment, and the second type of smoke candidate regions are selected. The intersection of the first type of smoke candidate region and the second type of smoke candidate region is selected as the confirmed smoke candidate region; Edge extraction and fusion are performed on the confirmed smoke candidate regions to obtain the location information of the smoke regions, thereby obtaining the final smoke detection result output; The segmentation of the smoke candidate region includes: 1) Based on the histogram information of the image, obtain the grayscale estimates of the sea surface and sky, and obtain the initial seed points for the sea surface and sky; The initial seed points for the sky and sea are the first and second extreme points of the histogram of the smoothed image; 2) Based on the statistical empirical value of the white highlight grayscale of the smoke region and the maximum grayscale value of the image, the initial seed points of the smoke are obtained; The initial seed point for the smoke is ;in, These are statistical empirical values ​​for smoke ash level. This represents the maximum grayscale value of the image. 3) Based on the initial seed points of the sea surface, sky, and smoke, a clustering algorithm is used to cluster the image grayscale, and the potential image regions belonging to the smoke target are segmented and extracted to obtain smoke candidate regions.

2. The optical smoke detection method combining prior knowledge and feature classification according to claim 1, characterized in that, Morphological processing is performed on the segmented smoke candidate regions, and target dilation is performed in the horizontal direction to close the segmented smoke candidate regions. Based on prior knowledge, including the shape and height of the smoke, candidate regions that do not meet the smoke characteristics are eliminated from closed smoke candidate regions. Connectivity labeling is performed on the removed smoke candidate regions to obtain the first type of smoke candidate regions.

3. The optical smoke detection method combining prior knowledge and feature classification according to claim 1, characterized in that, When screening candidate regions for the second type of smoke The multidimensional feature descriptor contains eight features: gray-level mean, gray-level variance, maximum gray-level value, gray-level difference mean, contrast, entropy, vertical gradient mean, and vertical gradient variance.

4. The optical smoke detection method combining prior knowledge and feature classification according to claim 3, characterized in that, The classifier uses a tree-based classification algorithm.

5. The optical smoke detection method combining prior knowledge and feature classification according to claim 1, characterized in that, The Kmeans gray-level clustering algorithm is used to cluster the gray levels of the image to obtain the potential regions of the image belonging to the smoke target. After segmentation and extraction, the smoke candidate regions are obtained.

6. The optical smoke detection method combining prior knowledge and feature classification according to claim 1, characterized in that, Methods for edge extraction and fusion of confirmed smoke candidate regions include: 1) An edge detection algorithm is used to extract the edge map of the smoke; 2) Based on the edge map information, the smoke is confirmed and the segmentation results are supplemented to obtain the segmentation map after edge fusion; 3) Perform morphological dilation on the segmented map after edge fusion to obtain the final edge feature fusion map; 4) Connected component labeling is performed on the edge feature fusion map to obtain the final smoke detection result output.

7. The optical smoke detection method combining prior knowledge and feature classification according to claim 1, characterized in that, For each confirmed smoke candidate region, the number of edge points in the corresponding position of the edge map is counted to determine its matching degree with the smoke candidate region. When the overlap rate condition is met, it is considered that the two have an intersection, and the current smoke candidate region is determined to be the smoke target region. The segmentation map after fusing edges is obtained by taking the union of the clustered smoke candidate regions and the edge maps corresponding to their positions.

8. The optical smoke detection method combining prior knowledge and feature classification according to claim 7, characterized in that, The segmented image after edge fusion: In the formula, For clustering segmentation graphs, For edge extraction map; For the first The edge map corresponding to the current smoke candidate region and the current smoke candidate region The matching degree of each smoke candidate region; For the first The edge map corresponding to the current smoke candidate region and the current smoke candidate region Overlap rate of candidate smoke regions; ; This represents the total number of smoke candidate regions in the cluster segmentation graph; These are the vertical and horizontal coordinates of the image; Specifically, the formula for calculating the matching degree is: The formula for calculating the overlap rate; in, Let be the coordinate information corresponding to the m-th smoke candidate region in the clustering segmentation map, represented by a rectangle, where , , , These represent the coordinates of the top, bottom, left, and right boundaries of the detection box, respectively. , Indicates the first The width and height of the rectangle; This is a configurable scaling factor.

Images