Target identification method based on artificial intelligence in complex environment

Abstract

The invention discloses a target recognition method and device based on artificial intelligence in a complex environment. The target recognition method comprises the steps of extracting image materials of a corresponding environment in a database based on a detection environment of a detection image; performing feature extraction on the image material to obtain a plurality of image feature contours in a corresponding environment; identifying and removing the image feature data of the complete contour in the detection image to obtain an incomplete image; extracting a plurality of incomplete feature contours from the incomplete image, and repairing the incomplete feature contours to obtain complemented image contour data; filling the plurality of repaired incomplete feature contours in the incomplete image, and verifying the contour data of the complemented image; and correcting the complemented image contour data based on the verification result, obtaining a plurality of feature contours in the detection image, and obtaining a target feature contour. By extracting the feature contour in the detection image and repairing and complementing the incomplete feature contour, the accuracy of target detection in the detection image is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a method for target recognition in a complex environment based on artificial intelligence. Background Art

[0002] Target recognition and extraction is one of the main methods of image processing. For pictures taken in complex scenes, the difficulty of identifying and extracting target units in the scene is increased because there are multiple interferences in the scene and multiple interferences block the target units. The existing recognition methods generally set multiple candidate frames and filter the areas where the target units appear to improve the accuracy of target recognition. Since there are many characteristic contours of interferences in complex scenes, there is interference with target detection in the candidate area, resulting in reduced accuracy of the candidate frame selection method. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art. The present invention provides a method and device for target recognition in a complex environment based on artificial intelligence, which improves the accuracy of target detection in the detection image by extracting the feature contours in the detection image and repairing and completing the incomplete feature contours.

[0004] The present invention provides a method for target recognition in a complex environment based on artificial intelligence, the target recognition method comprising:

[0005] Confirm the detection environment according to the detection image, and extract the image material of the corresponding environment from the database based on the detection environment;

[0006] Extract features from image materials to obtain several image feature contours in the corresponding environment;

[0007] Perform contour recognition on the detected image, identify and remove the image feature data of the complete contour, and obtain the incomplete image;

[0008] Extracting a plurality of incomplete feature contours from the incomplete image, repairing the incomplete feature contours according to the plurality of image feature contours, and obtaining completed image contour data;

[0009] Filling the incomplete image with the restored multiple incomplete feature contours, and verifying the contour data of the completed image;

[0010] The complement image contour data is corrected based on the verification result, a number of characteristic contours in the detection image are obtained, and a target characteristic contour is obtained from the several characteristic contours.

[0011] Furthermore, the step of confirming the detection environment according to the detection image and extracting image materials corresponding to the environment from the database based on the detection environment includes:

[0012] Input the detection image into the recognition system, obtain the detection environment according to the shooting environment of the detection image, and obtain the detection environment according to the shooting environment of the detection image;

[0013] According to the detection environment, image data of the same environment as the detection environment is extracted from a database, and the image data is extracted as image material.

[0014] Furthermore, the step of extracting features from the image material to obtain a plurality of image feature contours in the corresponding environment includes:

[0015] The contour of the image material is extracted by using the Canny algorithm, and different characteristic contours are extracted from the image material, so as to construct a characteristic contour database of the detection environment.

[0016] Furthermore, the extracting contours of the image material by using the Canny algorithm, extracting different characteristic contours from the image material, and thus constructing a characteristic contour database of the detection environment includes:

[0017] Performing gray value processing on the acquired image material, and performing Gaussian filtering on the gray value map of the acquired image material;

[0018] The Laplace transformed image is further sharpened by the non-maximum suppression method to obtain an accurate feature contour image.

[0019] Furthermore, the step of performing contour recognition on the detected image, identifying and removing image feature data of the complete contour, and obtaining the incomplete image includes:

[0020] Preprocessing the detection image to obtain a preprocessed image;

[0021] Perform contour recognition on the preprocessed image, mark and extract the complete feature contour;

[0022] The complete feature contour part in the preprocessed image is filled with background color to obtain an incomplete image.

[0023] Furthermore, the extracting of a plurality of incomplete feature contours from the incomplete image, and repairing the incomplete feature contours according to the plurality of image feature contours to obtain the completed image contour data comprises:

[0024] Extract the incomplete feature contour, and compare the incomplete feature contour with several image feature contours to obtain an approximate image feature contour;

[0025] The deflection angle of the approximate image feature is adjusted based on the incomplete feature contour, and the incomplete feature contour is completed based on the approximate image feature contour.

[0026] Furthermore, the extracting of the incomplete feature contour and obtaining an approximate image feature contour by comparing the incomplete feature contour with a plurality of image feature contours comprises:

[0027] An incomplete characteristic contour is acquired around several characteristic contours in the hollowed-out state in the incomplete image, and the incomplete contour feature is connected to the characteristic contour in the hollowed-out state.

[0028] Furthermore, the filling of the repaired plurality of incomplete feature contours in the incomplete image and the verification of the completed image contour data include:

[0029] Filling the completed image contour in the incomplete image, verifying the completed image contour based on the hollowed-out area in the incomplete image, and detecting whether the completed portion of the completed image contour is completely contained in the hollowed-out area of ​​the incomplete image;

[0030] If so, the completed image contour is marked as a repaired image contour; if not, the completed image contour is marked as an erroneous contour and the completed image contour is removed.

[0031] Further, the step of correcting the completed image contour data based on the verification result, obtaining a plurality of characteristic contours in the detection image, and obtaining a target characteristic contour from the plurality of characteristic contours includes:

[0032] Extracting a plurality of characteristic contours of the corrected detection image, and extracting a characteristic contour that matches the target characteristic contour comparison data among the plurality of characteristic contours, and marking them as target units;

[0033] The completed image contour data and the uncompleted incomplete feature contour data of the corrected detection image are obtained, and the first category target pending image data and the second category target pending image data are obtained by comparing the target feature contour control data.

[0034] The present invention also provides a target recognition device in a complex environment based on artificial intelligence, the recognition device comprising:

[0035] Image material acquisition module: used to confirm the detection environment according to the detection image, and extract the image material of the corresponding environment from the database based on the detection environment;

[0036] Feature extraction module: used to extract features from image materials and obtain several image feature contours in the corresponding environment;

[0037] Feature removal module: used to perform contour recognition on the detection image, identify and remove the image feature data of the complete contour, and obtain the incomplete image;

[0038] Feature repair module: used to extract several incomplete feature contours from the incomplete image, repair the incomplete feature contours according to several image feature contours, and obtain the completed image contour data;

[0039] Verification module: used to fill in the incomplete image with the restored several incomplete feature contours, and verify the contour data of the completed image;

[0040] Correction module: used to correct the completed image contour data based on the verification result, obtain several feature contours in the detection image, and obtain the target feature contour from several feature contours.

[0041] The present invention provides a method and device for target recognition in a complex environment based on artificial intelligence. According to the scene of a detection image, multiple image materials are extracted, and a complete feature contour in the detection image is removed to obtain an incomplete image. The incomplete feature contour in the incomplete image is completed based on the feature contours of the multiple image materials, thereby improving the convenience of extracting feature contours in the detection image and improving the accuracy of identifying target units in the detection image. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0043] Figure 1 is a flow chart of a method for target recognition in a complex environment based on artificial intelligence in an embodiment of the present invention;

[0044] Figure 2 It is a schematic diagram of a target recognition device in a complex environment based on artificial intelligence in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] Embodiment 1:

[0047] The embodiment of the present invention discloses a method for target recognition in a complex environment based on artificial intelligence. The detection environment is confirmed according to a detection image, and image materials of the corresponding environment are extracted from a database based on the detection environment. The detection environment corresponding to the detection image is confirmed by obtaining the shooting environment or site background of the detection image, so that image materials of the corresponding environment can be extracted from the database.

[0048] Feature extraction is performed on the image material to obtain several image feature contours in the corresponding environment. By obtaining the image feature contours in the image material, the feature contours of the detection image can be screened according to the image material, thereby improving the reliability of target detection.

[0049] The contour of the detection image is recognized, the image feature data of the complete contour is recognized and removed, an incomplete image is obtained, a number of incomplete feature contours are extracted from the incomplete image, the incomplete feature contours are repaired according to the several image feature contours, and the completed image contour data is obtained; by recognizing the feature contour of the detection image and removing the complete feature contour, the incomplete feature contour is repaired in combination with the image feature contour of the image material, and multiple feature contour information of the detection image can be obtained through repair, so as to detect the corresponding target contour features in the detection image.

[0050] The repaired several incomplete feature contours are filled in the incomplete image, the contour data of the completed image is verified, the contour data of the completed image is corrected based on the verification result, several feature contours in the detection image are obtained, and the target feature contour is obtained from the several feature contours. Each time the incomplete feature contour is repaired, the detection correction is performed based on the incomplete image formed by the detection image to ensure that the incomplete image repair can meet the repair requirements of the detection image and improve the reliability and accuracy of contour recognition of the detection image.

[0051] Figure 1 A flow chart of a method for identifying a target in a complex environment based on artificial intelligence in an embodiment of the present invention is shown. The method includes:

[0052] S11: Extract the detection environment.

[0053] Acquire the detection image and obtain the image data of the complex environment through the visual component, which can be an electronic device with shooting function such as a camera, a camera, and a mobile device.

[0054] The detection image is input into the recognition system, the detection environment is obtained according to the shooting environment of the detection image, the detection environment is obtained according to the shooting environment of the detection image, image data of the same environment as the detection environment is extracted from the database according to the detection environment, and the image data is extracted as image material.

[0055] Specifically, according to the detection shooting features of the detection image, the detection environment is extracted, and based on the keyword semantic retrieval, several image materials corresponding to the detection environment are obtained in the database. The selection of the image materials needs to meet the following rules: a single image material has fewer image features, and the image features of each image material are different from each other. By selecting image materials with fewer image features, the image feature extraction steps of the image materials can be simplified, and the convenience of feature contour extraction of the image materials can be improved; by setting several image materials with different image features, several image materials can include several different feature contours, thereby meeting the accuracy and reliability of the image materials in identifying the detection image contour features.

[0056] Furthermore, a plurality of the image materials may include image data corresponding to different shooting angles of the detection environment, thereby expanding the diversity of feature contours in the detection environment and improving the reliability of target detection and recognition.

[0057] S12: Extract feature contours.

[0058] The image material is subjected to contour extraction through the Canny algorithm, and different characteristic contours are extracted from the image material, so as to construct a characteristic contour database of the detection environment, so as to compare and identify the contour features of the detection image, so as to identify each contour feature of the detection image, thereby improving the connection reliability of obtaining the target characteristic contour in the detection image.

[0059] The Canny algorithm is a standard algorithm for edge detection. Its goal is to find an optimal edge detection solution or to find the location with the strongest change in grayscale intensity in an image, so as to accurately obtain the image feature contour.

[0060] The acquired image material is grayed out, and the gray value map of the acquired image material is Gaussian filtered. The gray value map is filtered by a Gaussian kernel. A pixel value is set on each grid on the gray value map. In this embodiment, a Gaussian kernel of 3*3 size is set so that the Gaussian kernel covers the gray value map. The pixel value after Gaussian convolution of the center grid corresponding to the Gaussian kernel is calculated. The pixel value of each grid of the gray value map is convoluted in turn by the Gaussian kernel. The gray value map is filtered by Gaussian kernel convolution, which can remove noise points of the gray value image and make the image of the gray value map smooth.

[0061] The gray value image after Gaussian filtering is subjected to partial derivative processing, and the location of the sudden change value of the image contour edge is confirmed by partial derivative, thereby improving the accuracy of contour extraction.

[0062] The partial derivative operation of the gray value image is performed based on the Laplacian operator. The Laplacian operator (Laplacian operator) is an isotropic edge extraction operator that can sharpen lines of any direction in the image to achieve a good smoothing effect.

[0063] The calculation formula of the Laplacian operator is:

[0064] Laplacian=f(x+1,y)+f(x-1,y)+f(x,y+1)+f(x,y-1)-4f(x,y);

[0065] Among them, Laplacian is the Laplace transform value, and f(x,y) is the coordinate value of the image pixel value.

[0066] The image after Laplace transformation is further sharpened by the non-maximum suppression method. According to the gray value data of the image edge, several candidate boxes are set, and the corresponding positions of the candidate boxes are set at the edge of the image. A local maximum value in the gray value of the image edge is searched, and the pixel values ​​of the candidate box corresponding to the local maximum value and the adjacent candidate box are calculated. The overlap rate between the adjacent candidate box and the maximum value candidate box is calculated, and the overlap rate is compared with a preset threshold. If the overlap rate is greater than the preset threshold, the corresponding candidate box is deleted, and only the candidate box with the maximum value is retained, so as to obtain the position of the most prominent change of the image edge, sharpen the image edge, and obtain an accurate feature contour image.

[0067] Furthermore, several characteristic contour images are extracted from the image material, and it is detected whether there is a target characteristic image contour in the several characteristic contour images. If so, the corresponding target characteristic image contour is copied and organized into target characteristic contour comparison data, and the target characteristic contour comparison data includes several target contour characteristic image data.

[0068] S13: Extract the complete image feature contour.

[0069] The contour of the detection image is recognized, and the image feature data of the complete contour is identified and removed to obtain the incomplete image. By obtaining the image contour involved in the detection image, the reliability of the detection image for contour recognition is improved.

[0070] Specifically, preprocessing the detection image to obtain a preprocessed image;

[0071] The detection image is grayed out to obtain a gray value map of the detection image so as to identify feature contours of the detection image.

[0072] The contour recognition is performed on the preprocessed image, and the contour features of the detection image are extracted through Gaussian filtering, Laplace transform and non-maximum suppression method.

[0073] A complete contour image is extracted from the contour image of the detection image by contour extraction, and the complete contour image is filled with background color, so that the complete feature contour in the detection image is in a hollow state, that is, the detection image is formed as an incomplete image.

[0074] An incomplete feature contour is obtained around several feature contours in the hollowed-out state in the incomplete image, and the incomplete contour feature is connected to the feature contour in the hollowed-out state, that is, an incomplete feature contour formed by a partially blocked feature contour is obtained, and by comparing with several feature contours extracted from the image material, the incomplete feature contour can be completed and repaired, so as to obtain an accurate feature contour in the detection image.

[0075] S14: Complete the incomplete feature contour.

[0076] The method extracts a plurality of incomplete feature contours from the incomplete image, and repairs the incomplete feature contours according to the plurality of image feature contours to obtain the completed image contour data. By extracting a plurality of image feature contours from the image material, and comparing and matching the image feature contours with the plurality of incomplete feature contours in the incomplete image, the image contour features that are successfully matched are repaired as the corresponding incomplete image feature contours, so as to repair the incomplete image and obtain the feature contour image in the incomplete image.

[0077] Extract the incomplete feature contour, and obtain an approximate image feature contour based on the comparison of the incomplete feature contour among several image feature contours, by extracting the incomplete feature contour at the edge of the hollowed-out area of ​​the incomplete image, and comparing several image feature contours of the image material with the incomplete feature contour, when the incomplete feature contour matches one or more of the several image feature contours, the successfully matched image feature contours are sorted into approximate image feature contours.

[0078] The deflection angle of the approximate image feature contour is adjusted based on the incomplete feature contour, and the incomplete feature contour is completed based on the approximate image feature contour. The deflection angle of the approximate image feature contour is adjusted so that the approximate image feature contour completely overlaps with the incomplete feature contour, thereby completing the incomplete feature contour.

[0079] Based on the image feature contour that successfully matches the incomplete feature contour, the incomplete feature contour is completed to obtain a plurality of completed image contour data.

[0080] S15: Correct feature contour.

[0081] Specifically, the completed image contour is filled in the incomplete image, and the completed image contour is verified based on the hollow area in the incomplete image to detect whether the completed part of the completed image contour is completely contained in the hollow area of ​​the incomplete image. If so, the completed image contour is marked as a repaired image contour; if not, the completed image contour is marked as an erroneous contour and the completed image contour is eliminated.

[0082] If the completed image contour is filled in the incomplete image, and the completed image contour can simultaneously meet the matching requirements of two incomplete feature contours, that is, the completed image contour forms two incomplete feature contours based on the hollow area, then the completed image contour is marked as a repaired image contour, and the remaining completed image contours corresponding to the two incomplete feature contours are eliminated.

[0083] Furthermore, the contour of the completed image is detected and corrected based on the hollowed-out area of ​​the incomplete image, and the contour of the repaired image obtained can meet the repair requirements of the incomplete image.

[0084] Furthermore, after the incomplete feature contour corresponding to the hollowed-out area in the incomplete image is completed and repaired, the repaired incomplete feature contour is filled with the background color, thereby expanding the hollowed-out area of ​​the incomplete image to obtain a new incomplete image, and the incomplete feature contour of the new incomplete image is completed and repaired until all the completed image feature contours of the detection image are obtained.

[0085] Specifically, the restoration image contour is matched with the target feature contour comparison data, and if the restoration image contour matches the target feature contour comparison data, the corresponding restoration image contour data is marked as the target feature contour data.

[0086] S16: Obtain target feature contour.

[0087] Specifically, a plurality of image feature contours are obtained in the detection image, and the image contour features in the detection image are repaired and completed in combination with the image contour features extracted from the image material, so as to obtain all completed image feature contours and uncompleted incomplete feature contours in the detection image.

[0088] A plurality of characteristic contours of the corrected detection image are extracted, and characteristic contours corresponding to and matching the target characteristic contour comparison data are extracted from the plurality of characteristic contours and marked as target units.

[0089] Compare the completed image feature contour in the detection image with the target feature contour comparison data one by one, obtain image data in the completed image feature contour that is similar to the target feature contour comparison data, and mark it as the first type of target pending image data;

[0090] The incomplete feature contour that has not been completed is matched with the target feature contour control data. If there is incomplete feature contour data that has been successfully matched, the incomplete feature contour data is marked as the second type of target pending image data.

[0091] According to the first category target pending image data and the second category target pending image data, the contour feature image of the corresponding position is marked in the detection image, and the target feature contour is output.

[0092] Specifically, the event probability that the first category target pending image data and the second category target pending image data are confirmed as target units is calculated. Since the number of target units existing in the detection image is a constant, the calculation coefficient of the first category target pending image data as the target unit is set to μ1, and the calculation coefficient of the second category target pending image data as the target unit is set to μ2, wherein μ1<μ2, and μ1+μ2=1.

[0093] Specifically, the distribution positions of the first-category target pending image data and the second-category target pending image data are obtained, and a preliminary weight is assigned to each of the first-category target pending image data and the second-category target pending image data according to the detection environment of the detection image and the characteristics of the target object.

[0094] Furthermore, the initial weight may be allocated based on the experience of detection-related staff, or may be allocated according to an initial weight allocation rule preset by the target recognition system.

[0095] Further, the probability that the first type of target to be determined image data is the target unit is calculated, and the probability calculation formula of the first type of target to be determined image data is the target unit is:

[0096]

[0097] Among them, σ1 is the standard deviation of the initial proportion of the target unit of the first type of target image data, a i is the initial weight of the i-th first-category target image data to be determined, μ is the average initial weight of the first-category target image data to be determined, N is the number of first-category target image data to be determined, CV i is the calculation coefficient of the undetermined image data of the i-th first-category target, is the probability that the i-th first-category target image data is a target unit.

[0098] Further, the probability that the second type of target to be determined image data is the target unit is calculated, and the probability calculation formula for the second type of target to be determined image data is the target unit is:

[0099]

[0100] Among them, σ2 is the standard deviation of the initial proportion of the second type of target undetermined image data as the target unit, b i is the initial weight of the i-th second-category target image data to be determined, μ′ is the average initial weight of the second-category target image data to be determined, M is the number of second-category target image data to be determined, CV′ i is the calculation coefficient of the undetermined image data of the i-th second-category target, is the probability that the i-th second-category target image data is a target unit.

[0101] By calculating the probability that each first-category target pending image and each second-category target pending image is a target unit, the mark of the target unit and the probability analysis can be output in the detection image, thereby improving the accuracy of target recognition detection in the detection image taken in a complex environment.

[0102] The embodiment of the present invention provides a target recognition method in a complex environment based on artificial intelligence. The detection environment of the detection image is identified, the image material of the corresponding detection environment is obtained, the contour features in the detection image are identified and extracted in combination with the image material, and the incomplete feature contours are repaired and completed after the incomplete image is obtained, so as to obtain multiple contour image features of the detection image, thereby improving the accuracy of detection and recognition of feature contours in the detection image, thereby improving the accuracy and convenience of identifying and outputting target contours in the detection image.

[0103] Embodiment 2:

[0104] Figure 2 A schematic diagram of a target recognition device based on artificial intelligence in a complex environment in an embodiment of the present invention is shown, and the target recognition device includes:

[0105] The image material acquisition module 10 is used to confirm the detection environment according to the detection image, and extract the image material of the corresponding environment from the database based on the detection environment.

[0106] The detection image is input into the recognition system, the detection environment is obtained according to the shooting environment of the detection image, the detection environment is obtained according to the shooting environment of the detection image, image data of the same environment as the detection environment is extracted from the database according to the detection environment, and the image data is extracted as image material.

[0107] Specifically, according to the detection shooting features of the detection image, the detection environment is extracted, and based on the keyword semantic retrieval, several image materials corresponding to the detection environment are obtained in the database. The selection of the image materials needs to meet the following rules: a single image material has fewer image features, and the image features of each image material are different from each other. By selecting image materials with fewer image features, the image feature extraction steps of the image materials can be simplified, and the convenience of feature contour extraction of the image materials can be improved; by setting several image materials with different image features, several image materials can include several different feature contours, thereby meeting the accuracy and reliability of the image materials in identifying the detection image contour features.

[0108] Feature extraction module 20: used to extract features from image materials and obtain a number of image feature contours in the corresponding environment.

[0109] The image material is subjected to contour extraction through the Canny algorithm, and different characteristic contours are extracted from the image material, so as to construct a characteristic contour database of the detection environment, so as to compare and identify the contour features of the detection image, so as to identify each contour feature of the detection image, thereby improving the connection reliability of obtaining the target characteristic contour in the detection image.

[0110] Feature removal module 30: used to perform contour recognition on the detection image, recognize and remove the image feature data of the complete contour, and obtain an incomplete image.

[0111] The contour of the detection image is recognized, and the image feature data of the complete contour is identified and removed to obtain the incomplete image. By obtaining the image contour involved in the detection image, the reliability of the detection image for contour recognition is improved.

[0112] Specifically, preprocessing the detection image to obtain a preprocessed image;

[0113] The detection image is grayed out to obtain a gray value map of the detection image so as to identify feature contours of the detection image.

[0114] The contour recognition is performed on the preprocessed image, and the contour features of the detection image are extracted through Gaussian filtering, Laplace transform and non-maximum suppression method.

[0115] A complete contour image is extracted from the contour image of the detection image by contour extraction, and the complete contour image is filled with background color, so that the complete feature contour in the detection image is in a hollow state, that is, the detection image is formed as an incomplete image.

[0116] An incomplete feature contour is obtained around several feature contours in the hollowed-out state in the incomplete image, and the incomplete contour feature is connected to the feature contour in the hollowed-out state, that is, an incomplete feature contour formed by a partially blocked feature contour is obtained, and by comparing with several feature contours extracted from the image material, the incomplete feature contour can be completed and repaired, so as to obtain an accurate feature contour in the detection image.

[0117] The feature repair module 40 is used to extract a plurality of incomplete feature contours from the incomplete image, and repair the incomplete feature contours according to a plurality of image feature contours to obtain completed image contour data.

[0118] The method extracts a plurality of incomplete feature contours from the incomplete image, and repairs the incomplete feature contours according to the plurality of image feature contours to obtain the completed image contour data. By extracting a plurality of image feature contours from the image material, and comparing and matching the image feature contours with the plurality of incomplete feature contours in the incomplete image, the image contour features that are successfully matched are repaired as the corresponding incomplete image feature contours, so as to repair the incomplete image and obtain the feature contour image in the incomplete image.

[0119] Extract the incomplete feature contour, and obtain an approximate image feature contour based on the comparison of the incomplete feature contour among several image feature contours, by extracting the incomplete feature contour at the edge of the hollowed-out area of ​​the incomplete image, and comparing several image feature contours of the image material with the incomplete feature contour, when the incomplete feature contour matches one or more of the several image feature contours, the successfully matched image feature contours are sorted into approximate image feature contours.

[0120] The deflection angle of the approximate image feature contour is adjusted based on the incomplete feature contour, and the incomplete feature contour is completed based on the approximate image feature contour. The deflection angle of the approximate image feature contour is adjusted so that the approximate image feature contour completely overlaps with the incomplete feature contour, thereby completing the incomplete feature contour.

[0121] Based on the image feature contour that successfully matches the incomplete feature contour, the incomplete feature contour is completed to obtain a plurality of completed image contour data.

[0122] Verification module 50: used to fill in the incomplete image with the restored incomplete feature contours, and verify the completed image contour data.

[0123] Specifically, the completed image contour is filled in the incomplete image, and the completed image contour is verified based on the hollow area in the incomplete image to detect whether the completed part of the completed image contour is completely contained in the hollow area of ​​the incomplete image. If so, the completed image contour is marked as a repaired image contour; if not, the completed image contour is marked as an erroneous contour and the completed image contour is eliminated.

[0124] If the completed image contour is filled in the incomplete image, and the completed image contour can simultaneously meet the matching requirements of two incomplete feature contours, that is, the completed image contour forms two incomplete feature contours based on the hollow area, then the completed image contour is marked as a repaired image contour, and the remaining completed image contours corresponding to the two incomplete feature contours are eliminated.

[0125] Correction module 60: used to correct the complement image contour data based on the verification result, obtain a number of characteristic contours in the detection image, and obtain a target characteristic contour from the several characteristic contours.

[0126] After the incomplete feature contour corresponding to the hollowed-out area in the incomplete image is completed and repaired, the repaired incomplete feature contour is filled with background color, so as to expand the hollowed-out area of ​​the incomplete image to obtain a new incomplete image, and the incomplete feature contour of the new incomplete image is completed and repaired until all the completed image feature contours of the detection image are obtained.

[0127] Specifically, the restoration image contour is matched with the target feature contour comparison data, and if the restoration image contour matches the target feature contour comparison data, the corresponding restoration image contour data is marked as the target feature contour data.

[0128] A plurality of characteristic contours of the corrected detection image are extracted, and characteristic contours corresponding to and matching the target characteristic contour comparison data are extracted from the plurality of characteristic contours and marked as target units.

[0129] Specifically, a plurality of image feature contours are obtained in the detection image, and the image contour features in the detection image are repaired and completed in combination with the image contour features extracted from the image material, so as to obtain all completed image feature contours and uncompleted incomplete feature contours in the detection image.

[0130] Compare the completed image feature contour in the detection image with the target feature contour comparison data one by one, obtain image data in the completed image feature contour that is similar to the target feature contour comparison data, and mark it as the first type of target pending image data;

[0131] The incomplete feature contour that has not been completed is matched with the target feature contour control data. If there is incomplete feature contour data that has been successfully matched, the incomplete feature contour data is marked as the second type of target pending image data.

[0132] According to the first category target pending image data and the second category target pending image data, the contour feature image of the corresponding position is marked in the detection image, and the target feature contour is output.

[0133] Specifically, the event probability that the first category target pending image data and the second category target pending image data are confirmed as target units is calculated. Since the number of target units existing in the detection image is a constant, the calculation coefficient of the first category target pending image data as the target unit is set to μ1, and the calculation coefficient of the second category target pending image data as the target unit is set to μ2, wherein μ1<μ2, and μ1+μ2=1.

[0134] Specifically, the distribution positions of the first-category target pending image data and the second-category target pending image data are obtained, and a preliminary weight is assigned to each of the first-category target pending image data and the second-category target pending image data according to the detection environment of the detection image and the characteristics of the target object.

[0135] Furthermore, the initial weight may be allocated based on the experience of detection-related staff, or may be allocated according to an initial weight allocation rule preset by the target recognition system.

[0136] Further, the probability that the first type of target to be determined image data is the target unit is calculated, and the probability calculation formula of the first type of target to be determined image data is the target unit is:

[0137]

[0138] Among them, σ1 is the standard deviation of the initial proportion of the target unit of the first type of target image data, a i is the initial weight of the i-th first-category target image data to be determined, μ is the average initial weight of the first-category target image data to be determined, N is the number of first-category target image data to be determined, CV i is the calculation coefficient of the undetermined image data of the i-th first-category target, is the probability that the i-th first-category target image data is a target unit.

[0139] Further, the probability that the second type of target to be determined image data is the target unit is calculated, and the probability calculation formula for the second type of target to be determined image data is the target unit is:

[0140]

[0141] Among them, σ2 is the standard deviation of the initial proportion of the second type of target undetermined image data as the target unit, b i is the initial weight of the i-th second-category target image data to be determined, μ′ is the average initial weight of the second-category target image data to be determined, M is the number of second-category target image data to be determined, CV′ i is the calculation coefficient of the undetermined image data of the i-th second-category target, is the probability that the i-th second-category target image data is a target unit.

[0142] By calculating the probability that each first-category target pending image and each second-category target pending image is a target unit, the mark of the target unit and the probability analysis can be output in the detection image, thereby improving the accuracy of target recognition detection in the detection image taken in a complex environment.

[0143] A person skilled in the art may understand that all or part of the steps in the various methods of the above embodiments may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the storage medium may include: a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, etc.

[0144] In addition, the embodiments of the present invention are described in detail above. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A method for target recognition in a complex environment based on artificial intelligence, characterized in that: The target recognition method comprises: Confirm the detection environment according to the detection image, and extract the image material of the corresponding environment from the database based on the detection environment; Extract features from image materials to obtain several image feature contours in the corresponding environment; Perform contour recognition on the detected image, identify and remove the image feature data of the complete contour, and obtain the incomplete image; Extracting a plurality of incomplete feature contours from the incomplete image, repairing the incomplete feature contours according to the plurality of image feature contours, and obtaining completed image contour data; Filling the incomplete image with the restored multiple incomplete feature contours, and verifying the contour data of the completed image; The complement image contour data is corrected based on the verification result, a number of characteristic contours in the detection image are obtained, and a target characteristic contour is obtained from the several characteristic contours.

2. The method for target recognition in a complex environment based on artificial intelligence as claimed in claim 1, characterized in that: The step of confirming the detection environment according to the detection image and extracting the image material of the corresponding environment from the database based on the detection environment includes: Input the detection image into the recognition system, obtain the detection environment according to the shooting environment of the detection image, and obtain the detection environment according to the shooting environment of the detection image; According to the detection environment, image data of the same environment as the detection environment is extracted from a database, and the image data is extracted as image material.

3. The method for target recognition in a complex environment based on artificial intelligence as claimed in claim 1, characterized in that: The step of extracting features from the image material to obtain a plurality of image feature contours in the corresponding environment includes: The contour of the image material is extracted by using the Canny algorithm, and different characteristic contours are extracted from the image material, so as to construct a characteristic contour database of the detection environment.

4. The method for target recognition in a complex environment based on artificial intelligence as claimed in claim 3, characterized in that: The method of extracting contours of the image material by using the Canny algorithm, extracting different characteristic contours from the image material, and thereby constructing a characteristic contour database of the detection environment includes: Performing gray value processing on the acquired image material, and performing Gaussian filtering on the gray value map of the acquired image material; The Laplace transformed image is further sharpened by the non-maximum suppression method to obtain an accurate feature contour image.

5. The method for target recognition in a complex environment based on artificial intelligence as claimed in claim 1, characterized in that: The method of performing contour recognition on the detection image, identifying and removing image feature data of the complete contour, and obtaining the incomplete image comprises: Preprocessing the detection image to obtain a preprocessed image; Perform contour recognition on the preprocessed image, mark and extract the complete feature contour; The complete feature contour part in the preprocessed image is filled with background color to obtain an incomplete image.

6. The method for target recognition in a complex environment based on artificial intelligence as claimed in claim 1, characterized in that: The extracting of a plurality of incomplete feature contours from the incomplete image, and repairing the incomplete feature contours according to the plurality of image feature contours to obtain the completed image contour data comprises: Extract the incomplete feature contour, and obtain an approximate image feature contour by comparing the incomplete feature contour with several image feature contours; The deflection angle of the approximate image feature is adjusted based on the incomplete feature contour, and the incomplete feature contour is completed based on the approximate image feature contour.

7. The method for target recognition in a complex environment based on artificial intelligence as claimed in claim 6, characterized in that: The extracting of the incomplete feature contour and obtaining an approximate image feature contour by comparing the incomplete feature contour with a plurality of image feature contours comprises: An incomplete characteristic contour is acquired around several characteristic contours in the hollowed-out state in the incomplete image, and the incomplete contour feature is connected to the characteristic contour in the hollowed-out state.

8. The method for target recognition in a complex environment based on artificial intelligence as claimed in claim 1, characterized in that: The filling of the repaired plurality of incomplete feature contours in the incomplete image and the verification of the completed image contour data include: Filling the completed image contour in the incomplete image, verifying the completed image contour based on the hollowed-out area in the incomplete image, and detecting whether the completed portion of the completed image contour is completely contained in the hollowed-out area of ​​the incomplete image; If so, the completed image contour is marked as a repaired image contour; if not, the completed image contour is marked as an erroneous contour and the completed image contour is removed.

9. The method for target recognition in a complex environment based on artificial intelligence as claimed in claim 1, characterized in that: The method of correcting the complement image contour data based on the verification result, obtaining a plurality of characteristic contours in the detection image, and obtaining a target characteristic contour from the plurality of characteristic contours includes: Extracting a plurality of characteristic contours of the corrected detection image, and extracting a characteristic contour corresponding to and matching the target characteristic contour comparison data from the plurality of characteristic contours, and marking them as target units; The completed image contour data and the uncompleted incomplete feature contour data of the corrected detection image are obtained, and the first category target pending image data and the second category target pending image data are obtained by comparing the target feature contour control data.

10. A target recognition device based on artificial intelligence in a complex environment, characterized in that: The identification device comprises: Image material acquisition module: used to confirm the detection environment according to the detection image, and extract the image material of the corresponding environment from the database based on the detection environment; Feature extraction module: used to extract features from image materials and obtain several image feature contours in the corresponding environment; Feature removal module: used to perform contour recognition on the detection image, identify and remove the image feature data of the complete contour, and obtain the incomplete image; Feature repair module: used to extract several incomplete feature contours from the incomplete image, repair the incomplete feature contours according to several image feature contours, and obtain the completed image contour data; Verification module: used to fill in the incomplete image with the restored several incomplete feature contours, and verify the contour data of the completed image; Correction module: used to correct the completed image contour data based on the verification result, obtain several feature contours in the detection image, and obtain the target feature contour from several feature contours.

Images