WO2022001843A1 - 图像特征提取方法和设备 - Google Patents

图像特征提取方法和设备 Download PDF

Info

Publication number
WO2022001843A1
WO2022001843A1 PCT/CN2021/102237 CN2021102237W WO2022001843A1 WO 2022001843 A1 WO2022001843 A1 WO 2022001843A1 CN 2021102237 W CN2021102237 W CN 2021102237W WO 2022001843 A1 WO2022001843 A1 WO 2022001843A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
feature
feature extraction
region
regions
Prior art date
Application number
PCT/CN2021/102237
Other languages
English (en)
French (fr)
Inventor
童强
许宽宏
Original Assignee
索尼集团公司
童强
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 索尼集团公司, 童强 filed Critical 索尼集团公司
Priority to CN202180046722.9A priority Critical patent/CN115956260A/zh
Publication of WO2022001843A1 publication Critical patent/WO2022001843A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components

Definitions

  • the present disclosure relates to image feature extraction.
  • object detection/recognition/comparison/tracking in still images or a series of moving images is widely and importantly applied to and plays an important role in the fields of image processing, computer vision and recognition.
  • Objects may be human body parts, such as face, hands, body, etc., other living things or plants, or any other object that is desired to be detected.
  • Object recognition/verification is one of the most important computer vision tasks, where the goal is to accurately identify or verify a specific object in it based on input photos/videos.
  • image features/object features are often obtained, and then object recognition/verification is performed based on the obtained image/object features.
  • images obtained by current camera devices often contain various noises, and the presence of noises deteriorates image quality, which may result in inability to extract sufficiently accurate image features/object features, which in turn affects object recognition.
  • the solution of the present disclosure can suppress or eliminate the influence of noise and extract more accurate image features by extracting the features of the image regions contained in the image and coding and combining the extracted features of the image regions.
  • an image feature extraction apparatus including a processing circuit, the processing circuit may be configured to: divide an image into at least two image areas; divide the image area of each of the at least two image areas The image region feature value is compared with the specific feature value to set a feature indicator of the image region based on the comparison result; and the feature indicator of each image region is combined to determine a feature of the image.
  • an image feature extraction method may include dividing an image into at least two image regions; performing a comparison to set feature indicators of the image regions based on the results of the comparison; and combining the feature indicators of the respective image regions to determine a feature of the image.
  • a method comprising at least one processor and at least one storage device having stored thereon instructions that, when executed by the at least one processor, cause the at least one processor to A processor performs methods as described herein.
  • a storage medium storing instructions that, when executed by a processor, can cause performance of a method as described herein.
  • an apparatus in yet another aspect, includes means for performing a method as described herein.
  • Figure 1 shows a general image feature extraction operation flow.
  • FIG. 2 shows a flowchart of an image feature extraction method according to the present disclosure.
  • FIG. 3 shows an image feature extraction process according to a preferred embodiment of the present disclosure.
  • FIG. 4 shows a block diagram of an image feature extraction apparatus according to the present disclosure.
  • FIG. 5 shows an exemplary image feature extraction process according to the first embodiment of the present disclosure.
  • FIG. 6 shows an intermediate output of an image feature extraction process according to the first embodiment of the present disclosure.
  • FIG. 7 illustrates an image area around a keypoint according to an embodiment of the present disclosure.
  • FIG. 8 shows an image feature extraction process according to the second embodiment of the present disclosure.
  • FIG. 9 shows the intermediate output of the image feature extraction process according to the second embodiment of the present disclosure.
  • FIG. 10 shows a block diagram illustrating an exemplary hardware configuration of a computer system capable of implementing embodiments of the present invention.
  • image feature extraction is widely used in various applications, such as object recognition, detection, verification, alignment, tracking, etc., and plays an important role.
  • an object may be a human body part, such as a face, hand, body, etc., other living beings or plants, or any other object that is desired to be detected.
  • it is necessary to perform feature extraction on an image, and then perform various subsequent processing/operations based on the extracted features.
  • a general image feature extraction operation flow will be described below with reference to FIG. 1 .
  • Image acquisition is performed to obtain an image containing the object to be recognized/verified.
  • Image acquisition can be achieved by various known image acquisition devices, such as cameras, cameras, drones, etc., which will not be described in detail here.
  • an image may refer to any of a variety of images, such as color images, grayscale images, and the like. It should be noted that in the context of this specification, the type of image is not particularly limited as long as such an image can be subjected to processing in order to extract image/object features therefrom, which in turn can be used to identify/detect objects in the image.
  • the image may be any suitable image, such as a raw image obtained by a camera, or an image that has undergone specific processing/preprocessing on the raw image, such as preliminary filtering, antialiasing, distortion removal, alignment, color adjustment, Contrast adjustment, normalization, and more. It should be noted that this particular process may also include other types of preprocessing operations known in the art, which will not be described in detail here. It should be noted that the advantageous technical effects of the present application can still be achieved even without the above-mentioned specific processing.
  • an image feature is extracted for the obtained image, which may be, in particular, an object feature of an object to be recognized contained in the image, and then the extracted object feature is output for subsequent operations.
  • an image containing an object refers to an image of an object whose image contains the object.
  • the object image may also sometimes be referred to as the object region in the image.
  • Object recognition also refers to the recognition of images in the object area. Therefore, in the context of the present disclosure, especially in the case of object recognition/verification/tracking, etc., it is necessary to accurately acquire object features for object recognition/verification/tracking, etc.
  • image features often refer to the to-be-identified /Verify object features of objects
  • image feature extraction can often refer to object feature extraction of objects to be identified/verified/tracked, unless otherwise stated.
  • the object area/object image in the image may be acquired first, and then feature extraction is performed on the acquired object area/object image.
  • the captured image may first be processed to obtain an image of the object in it, such as segmenting the image, etc., and then mainly feature the object area/object image with the object as the main body extract.
  • Image features/object features can be represented in various suitable ways.
  • features of an object particularly representative features
  • feature vector For example, in the case of detecting a face, pixel texture information, position coordinates, and the like of a representative part of the face are selected as features to form a feature vector of the image. Thereby, based on the obtained feature vector, object recognition/detection/tracking can be performed.
  • the feature vector may vary according to the model used in object recognition, and is not particularly limited.
  • the image feature/object feature may also relate to other attributes, characteristics of the image or be expressed in other ways, but is not limited thereto.
  • object recognition refers to determining whether the object in the image to be recognized is a specific object according to the extracted image features for an input image
  • object authentication/verification refers to at least two input images, according to The similarity between the extracted image features of at least two images is used to determine whether the objects in each image are the same object.
  • High-quality image features/object features are very important for subsequent processing results.
  • the images acquired by the current high-resolution image acquisition devices are often accompanied by a large amount of noise while capturing a large amount of details.
  • a large amount of noise information in the image area seriously interferes with the extraction of image features, resulting in that the existing feature extraction methods cannot extract features or the extracted features cannot effectively characterize the image, thus seriously affecting the subsequent recognition and classification based on image features.
  • Work That is to say, existing feature extraction methods cannot extract accurate/high-quality features through image pixels, for example, cannot directly extract image features with strong anti-noise and/or strong image representation capabilities.
  • the identification and classification of images such as fingerprints, faces, and human irises are widely used in the field of security. It is necessary to provide high-quality features to ensure high classification and identification accuracy.
  • local textures such as fingerprints, faces, and human iris
  • the existing feature extraction methods such as scale-invariant feature transform (SIFT), accelerated robust feature (SURF), etc. Accurate/high-quality features cannot be successfully extracted from images such as high-resolution fingerprints, which will adversely affect the classification and recognition accuracy.
  • SIFT scale-invariant feature transform
  • SURF accelerated robust feature
  • the present disclosure proposes an improved image feature extraction scheme.
  • the present disclosure proposes to extract features of an image region contained in an image, and to determine the features of the image based on a relative relationship between image region feature values of the image region and specific feature values.
  • the image noise may change the eigenvalues of pixels in a certain local area, but the relative relationship between the regional eigenvalues can be appropriately reduced or even eliminated.
  • Influence of noise in particular, by using the relative relationship to characterize image features, the influence of noise on the local image can be effectively reduced or even eliminated, so that the extracted image features have strong noise resistance and can more accurately characterize the image.
  • FIG. 2 shows a flowchart of an image feature extraction apparatus according to an embodiment of the present disclosure.
  • the image is divided into at least two image areas.
  • the image may be an original image or a preprocessed image.
  • the image may in particular be an image of an object, for example, may be an image of the object to be identified.
  • an image may be appropriately divided to obtain at least two image regions.
  • the image area may comprise at least one pixel, preferably an appropriate number of pixels according to the shape of the image area.
  • the image areas may be divided in any manner as long as the entire image can be covered, especially covering the entire image while being substantially adjacent to each other and not overlapping.
  • the shapes of the divided image regions are the same.
  • the image area may be any suitable shape, such as a square, a regular pentagon, a regular hexagon, a rectangle, and the like.
  • the determination of the image area may also depend on the extraction method of the image area feature, which will be described in detail below.
  • step S202 the image region feature value of each of the at least two image regions is compared with a specific feature value to set a feature indicator of the image region based on the comparison result.
  • the features of the image regions may be extracted in various known ways.
  • the image region features can be extracted according to a model that conforms to a function of Gaussian distribution, so as to obtain relatively robust features.
  • the image region features can be extracted according to a model that obeys Gauss-Hermit moments.
  • image region features may be represented in various ways.
  • a vector can be used, where each element in the vector corresponds to the value of a feature element.
  • the image region feature may be represented by a specific combined value of the values of each feature element, such as a statistical value, a weighted statistical value, and the like.
  • image area features can be represented by more compact values, improving storage and transmission efficiency, etc.
  • the feature value of the image area may be a mathematical statistical value of the feature of each pixel included in the image area.
  • the characteristics of an image area are determined based on Gaussian function-based statistical characteristics of pixels in the image area.
  • the features of an image region are determined based on Gauss-Hermit moment features of pixels in the image region.
  • the specific characteristic value may be set in an appropriate manner.
  • specific feature values may be preset, eg, based on previous training, detection results, eg, feature values related to the category of the object to be recognized, and the like.
  • the specific feature value may be related to a feature of the image region.
  • the specific feature is related to a feature of the selected reference image region of the at least two image regions.
  • the specific characteristic value is related to a characteristic of at least one of the at least two image areas, such as its characteristic or a mathematical statistical value of the characteristic.
  • the specific feature value may be applied to all image regions, which may be preset, or obtained by statistical calculation according to feature values of all image regions, such as mean value, median value, and the like.
  • the specific feature value may vary from image region to region, eg for an image region the specific feature value is the feature value of at least one adjacent region of which it is in a specific orientation. For example, in the case of grid-like segmentation, at least one of the eight image areas surrounding the image area may serve as its adjacent area.
  • the specific eigenvalue may be the eigenvalue of the adjacent area; if there are more than one adjacent area, the specific eigenvalue may be the mathematical value of the eigenvalues of the plurality of adjacent areas Statistical values such as mean, median, etc.
  • the feature value of each image region can be compared with a specific feature value corresponding to the image region, thereby obtaining the feature of the image region based on the comparison result.
  • the image region feature and the specific feature value are expressed in the same manner.
  • certain feature values may have the same vector form, ie have the same number of vector elements, contain the same class of vector elements, and so on.
  • the comparison between the image region feature and the specific feature value can be performed corresponding to the corresponding vector elements in the vector.
  • the comparison result indicates a relative relationship between the image region feature and the specific feature value.
  • the feature indicator can be represented in an appropriate form depending on the representation of the feature value/comparison result.
  • the feature indicator may be in the form of a vector, or a statistical value obtained by performing statistics on vector element values.
  • each element in the vector is the result of a comparison between the image region feature and the corresponding element of the specific feature value, and the comparison result indicates the relative relationship between the two elements.
  • the feature indicator may be represented in various ways, preferably in binary form.
  • a feature indicator can represent in binary form whether a region's feature is better than a particular feature value, for example, it can be 1 for better and 0 for no better.
  • the feature indicator may use a ratio to indicate the relative relationship between the feature of the region and the specific feature value, for example, may indicate the ratio/ratio between the feature of the region and the specific feature value, such as each The ratio/ratio between elements, or the ratio/ratio between the statistics of the feature. It should be noted that the comparison may correspond to one encoding, and that other ways of generating relative relationships are also applicable.
  • step S203 the feature indicators of each image region are combined to determine the feature of the image.
  • the feature of the image is a feature in the form of a vector consisting of feature indicators of individual image regions.
  • a higher-dimensional image feature can be constructed by combining the features of the image region through coding.
  • the combination mode is an encoding combination mode based on local binary comparison.
  • the comparison result ie, the feature indicators of the image regions
  • the comparison result is a binary representation
  • the features of the image are obtained by concatenating the feature indicators of the respective image regions.
  • the feature values of the image regions are in the form of vectors and the feature indicators of the image regions are in the form of corresponding binary vectors. In this way, performing a binary comparison between image regions and then performing a coding combination can improve the image representation capability of the anti-noise binary comparison result.
  • image area determination according to a preferred embodiment of the present disclosure will be described below, wherein the image area determination may also depend on the way of extracting image features.
  • the image area may be divided according to the distribution of key points of the image.
  • a keypoint refers to a corresponding location in an image that indicates a representative feature of the image.
  • the positions corresponding to possible human facial features and some key positions representing the posture of the face can be called key points.
  • the key points may indicate other representative locations in the image, which will not be described in detail here.
  • the determined key points can be to divide the image.
  • the image region obtained by this division can more appropriately reflect the feature distribution of the image, and then the image region features can more appropriately characterize the image.
  • FIG. 3 shows an image feature extraction process 300 according to an embodiment of the present disclosure.
  • step S301 the key point information of the image is determined, including the position of the key point and the scale of the key point.
  • key points may be obtained in various suitable ways.
  • the keypoints are computed from at least one underlying sub-feature of the image.
  • the multiple basic sub-features may be multiple statistical features based on a Gaussian function.
  • the plurality of basic sub-features are Gauss-Hermit moment features.
  • step S302 the region division of the image is determined according to the determined keypoint information and the keypoint scale.
  • the image area may be a local area around the detected keypoint in the image and having a size based on the detected keypoint scale information.
  • the image area may be an area in the image surrounding the detected keypoint, the keypoint may be the center point of the image area, and the size of the local area corresponds to the keypoint scale, such as a certain multiple of the keypoint scale .
  • Image regions may correspond to keypoints.
  • each keypoint may correspond to an image region.
  • any one of the two or more keypoints can be selected as the basis for Build the image area.
  • the image region may be a sub-region contained in a local region surrounding the detected keypoint in the image and having a size based on the detected keypoint scale information.
  • the image area is a sub-area obtained by further dividing the image local area determined based on the key points.
  • the division of the sub-regions may be performed in any suitable manner, as long as it can properly cover the partial regions.
  • the sub-region division may be in the form of a simple grid, etc., which will not be described in detail here.
  • the characteristics of the sub-regions may be determined in any suitable manner, such as those described above, which will not be described in detail here.
  • the features of the divided image regions may be acquired.
  • the features of the image area can be acquired in various suitable ways.
  • a model that follows a Gaussian distribution is used for feature extraction, so that the statistical features based on the Gaussian distribution have both a certain degree of anti-noise ability and image representation ability.
  • the characteristics of the image area may be determined in dependence on the determination of keypoints.
  • the features of the image area can be obtained in the same way as the keypoint determination. For example, it is obtained using the same or similar template as that used to obtain the keypoints.
  • keypoint detection is performed based on a Gaussian function, and wherein the characteristics of the image area are determined based on Gaussian function-based statistical characteristics of pixels in the image area.
  • keypoint detection is performed based on Gauss-Hermit moments, and wherein the features of the image region are determined based on Gauss-Hermit moments features of pixels in the image region. It should be noted that the above-mentioned acquisition of the features of the image area may be applicable to the case where the image area corresponds to a local area and a sub-area of the local area and so on.
  • step S304 the image region feature value of each of the at least two image regions is compared with the specific feature value to set the feature indicator of the image region based on the comparison result; and in step S305, the respective image regions are combined Characteristic indicators of the image area to determine the characteristics of the image.
  • the same comparison rules can be used for all image regions, for example, the features of all image regions are compared with the same specific feature value, or the features of each image region are compared with the features of adjacent regions at the same relative orientation .
  • image regions can be simply paired during the comparison process, and then every two image regions are compared. For example, two image regions that are adjacent to each other in a particular direction can be paired and compared.
  • the comparison result is thus obtained as a characteristic indicator of the image area.
  • the characteristic indicator may be a binary representation, or any other suitable relative ratio representation, which will not be described in detail here.
  • a sub-area feature indicator in the case where the image area corresponds to a sub-area of a local area, can be constructed by comparing the features of the sub-areas, so that the feature of the local area including the sub-area can be obtained, and then the sub-area feature indicator can be obtained.
  • Image features are obtained based on features of local regions.
  • the sub-region feature comparison can also be performed in various suitable manners, in particular in a binary or ratio manner as described above, which will not be described in detail here. Further sub-region division and comparison can better obtain the detailed information of the local image region, that is, the relationship between each sub-region and other sub-regions is used to express the part of the local image region that is different from other local image regions. In this way, The noise in the local image area to which each key point belongs can also be suppressed or eliminated by using the relative relationship, and the difference between each local image area and other local image areas can be distinguished by the coding relationship among all sub-areas within it.
  • the feature indicators of all image regions are combined to obtain a representation of the entire image feature.
  • the entire image feature may be in the form of a vector, where each element corresponds to a feature indicator for an image region.
  • the feature indicators of the image region may also be in the form of vectors, and then the resulting entire image features may be in the form of a matrix, in which a row or a column corresponds to the feature indicators of an image region vector of characters.
  • the image key point determination and feature extraction are carried out in an appropriate way, especially the key point determination and feature extraction based on the Gauss-Hermit moment, so as to obtain the basic sub-features with certain noise resistance and image representation ability.
  • the feature comparison is performed to further improve the noise resistance and image representation ability of the finally obtained image features.
  • FIG. 4 shows a block diagram of an electronic device for image feature extraction according to an embodiment of the present disclosure.
  • the electronic device 400 includes a processing circuit 420 that can be configured to divide the image into at least two image regions; compare the image region feature value of each of the at least two image regions with a specific feature value , to set the feature indicators of the image regions based on the comparison results; and combining the feature indicators of the respective image regions to determine the features of the image.
  • the processing circuit 420 may be in the form of a general-purpose processor, or may be a special-purpose processing circuit, such as an ASIC.
  • the processing circuit 420 can be constructed of a circuit (hardware) or a central processing device such as a central processing unit (CPU).
  • a program (software) for operating the circuit (hardware) or a central processing device may be carried on the processing circuit 420 .
  • the program can be stored in a memory such as arranged in a memory or an external storage medium connected from the outside, and downloaded via a network such as the Internet.
  • the processing circuit 420 may include various units for implementing the above-mentioned functions, in particular, the processing circuit 420 may include a dividing unit 422 configured to divide an image into at least two image regions; a comparing unit 424 for comparing the image area feature value of each of the at least two image areas with a specific feature value to set a feature indicator of the image area based on the comparison result; and being configured to combine The feature indicators for each image region determine a combining unit 426 of features of the image. Further, the processing circuit 420 may further include a keypoint determination unit 428 configured to determine keypoint information of the image (including keypoint positions and keypoint scales) and a feature extraction unit 430 configured to extract image region features. Each unit may operate as described above and will not be described in detail here.
  • the processing circuit 420 may further include a storage unit 432 for storing various information generated by the processing circuit 420 and other information required for processing operations.
  • the storage unit may be volatile memory and/or non-volatile memory.
  • memory may include, but is not limited to, random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), read only memory (ROM), flash memory.
  • the key point determination unit 428 and the feature extraction unit 430 may be located in the terminal-side electronic device but outside the processing circuit, and may even be located outside the electronic device 400 .
  • the electronic device of the present application can still work effectively and obtain the above-mentioned advantageous technical effects.
  • the memory unit may also be located within the electronic device but outside the processing circuitry, or even outside the electronic device. It should be noted that although each unit is shown as a separate unit in FIG. 4 , one or more of these units may also be combined into one unit, or split into multiple units.
  • each of the above-mentioned units may be implemented as independent physical entities, or may also be implemented by a single entity (eg, a processor (CPU or DSP, etc.), an integrated circuit, etc.).
  • the above-mentioned respective units are shown with dotted lines in the drawings to indicate that these units may not actually exist, and the operations/functions implemented by them may be implemented by the processing circuit itself.
  • FIG. 4 is only a schematic structural configuration of an electronic device for image feature extraction, and the electronic device 400 may also include other possible components.
  • the electronic device 400 may also include other components not shown, such as a network interface, a controller, and the like.
  • the processing circuitry may be associated with external memory.
  • the processing circuitry may be connected directly or indirectly (eg, with other components interposed therebetween) to external memory for accessing data.
  • FIG. 5 shows the image feature extraction process according to the first embodiment of the present application.
  • the determination of image key points and the extraction of image features are carried out sequentially.
  • the key point determination of the image and the extraction of image area features are performed based on the Gauss-Hermit moment.
  • step S501 the image is convolved with the first set of filter templates to determine the position and scale of key points in the image response map.
  • the first group of filter templates mainly refers to the image processing functions used to determine the key point information.
  • the first set of filter templates consists of a certain number of filter templates, eg, a certain number of filter templates of the same type but with different parameters (eg, size, order, etc.).
  • the type and number of templates may be appropriately set, for example, may be set in consideration of at least one of calculation time, calculation amount, calculation complexity, precision, recognizability, etc., so as to be able to Critical point determination is achieved cost-effectively.
  • the first set of filter templates contains a certain number of filter templates of variable size.
  • the first set of filter templates is used to obtain a template response map to obtain keypoint positions. where the dimensions of all filter templates are the same.
  • the specific size may be set empirically, for example, may be an empirical value related to the category of the object to be recognized, or may be determined in consideration of at least one of computational accuracy, recognizability, and the like Settings, for example, as described above.
  • the template response refers to the response to each pixel, for example, a feature value is obtained for each pixel, and then key points are selected by comparing and determining the feature values of the pixels.
  • the scale of the keypoints is determined by iteratively changing the size of the filter template.
  • the scale of key points is determined by finding extreme points in the image scale space composed of response maps obtained by multiple variable-size filter templates.
  • the number of iterations and the size change step size in each iteration may be set empirically, or may be set in consideration of at least one of computational accuracy, recognizability, etc., for example, as previously described.
  • Gauss-Hermit moments are defined as shown in (1) and (2) below.
  • GHM pq is the Gaussian calculated from the image area with width 2w+1 and height 2h+1 using a p+q order Gauss-Hermit moment filter template with width 2w+1 and height 2h+1 -
  • the response value of the Hermitian moment, f(x,y) is the gray value of the image pixel at the x,y position of the image, H p (x; ⁇ ), H q (y; ⁇ ) is such as (2 ), ⁇ indicates the size/dimension of the Gaussian-Hermit moment filter template, p, q indicate the order of the Hermite function.
  • the Gauss-Hermit moment has both the noise suppression ability of the Gaussian function and the image representation ability of the Hermite function. In the present disclosure, it can be used as an ideal basic sub-feature to replace the pixel grayscale that is susceptible to noise and has insufficient image representation ability.
  • the order of the Gauss-Hermit moment can be appropriately set, for example, set according to experience, or especially considering the performance indicators such as calculation amount, calculation complexity, identifiability, etc. is set as a compromise.
  • the performance indicators such as calculation amount, calculation complexity, identifiability, etc.
  • Higher-order Gauss-Hermit moments such as more than 4-order Gaussian-Hermit moments, have higher image representation capabilities, but their computational complexity is significantly increased and the ability to suppress noise is weakened.
  • image keypoint detection is performed using a filter template having at least one Gauss-Hermitian moment of variable size as a first set of filter templates.
  • the first set of filter templates may include variable size GHM 20 , GHM 02 , GHM 11 templates. Its ⁇ is preferably a multiple of 1.02. It should be noted that other variable size templates are also possible.
  • the Gauss-Hermit moment response map obtained by convolution of multiple variable-size GHM 20 , GHM 02 , GHM 11 templates is used for image keypoint detection, as shown in response map 1 and response map 2 in Figure 6 and the response shown in Figure 3.
  • any order of Gauss-Hermit moments or a combination of Gauss-Hermit moments can be used for keypoint detection. But in this embodiment, it is preferable to use the function composed of GHM 20 , GHM 02 , and GHM 11 of the same size template as the key point detection function.
  • GHM 1 GHM 10 2 +GHM 01 2 (5)
  • a plurality of a plurality of variable size GHM 20, GHM 02, GHM 11 Gaussian convolution template - Hemet scale space within the image in response to moments FIG composition is preferably by looking for (3 ) to determine the scale of key points, so that the feature extraction process is highly robust to image scale changes.
  • the number of response maps and the number of templates here is merely exemplary, and other numbers of templates and response maps are possible, and generally the number of response maps corresponds to the number of templates used.
  • the function shown has the ability to be invariant to the image rotation and has a considerable degree of noise resistance and image representation ability, and can detect enough unique key points from the image.
  • other Gauss-Hermit moment combination functions such as those shown in (4) or (5) can also be used to detect keypoints, but considering the robustness to image rotation changes or from the image
  • the number of detected key points and the influence of computational complexity are preferably the function shown in (3).
  • first determining the position of the key point and then determining the scale of the key point are only exemplary, and preferably, it can also be determined simultaneously by determining the extreme value of the function after the response map space is obtained by using the filter template of variable size.
  • Keypoint information especially keypoint location and keypoint scale.
  • other methods for obtaining key points or key point scales can also be applied to the embodiments of the present disclosure, which will not be described in detail here.
  • step S502 the image is convolved with the second set of filter templates to obtain an image response map.
  • each filter template in the second set of filter templates is a fixed size template.
  • the second set of filter templates may be identical, partially identical, or completely different from the first set of filter templates.
  • the features of the image regions are also acquired in the same way as the keypoints are determined.
  • each filter template in the second set of filter templates may have the same size but different orders.
  • the order of the Gauss-Hermit moment and the size of the filter template can have any combination, in this embodiment, it is preferable to use templates of 10 types of Gauss-Hermit moments as the second set of filters Templates, including fixed-size GHM 00 , GHM 10 , GHM 01 , GHM 30 , GHM 03 , GHM 21 , GHM 12 , a total of 7 templates, and the aforementioned first group of filter templates GHM 20 , GHM 02 , GHM 11 templates.
  • the order of the filter template can be set as described above, and Gauss-Hermit moments with orders from 0 to 3 are preferably used in this method.
  • the size of the filter templates can be the same, ⁇ is preferably 1.02.
  • the response graphs of the images can be obtained from these 10 templates, as shown in the response graphs n1, n2, n3, . . . , n9 in FIG. 6 .
  • the number of response maps n1, n2, n3, ..., n9 is only an example, it can also contain any other number of response maps, which number can be preset, or after considering noise immunity, computational complexity It is set in the case of a balance with the ability to express the image.
  • GHM 00 GHM 10 , GHM 01 , GHM 30 , GHM 03 , GHM 21 , GHM 12 , and GHM 20 , GHM 02 having the same ⁇ as the above-mentioned GHM 00 etc.
  • GHM 11 convolves a total of 10 templates to obtain a response map group of Gauss-Hermit moments, including 10 response maps corresponding to each template one-to-one, and then generates a new response map by combining internal functions
  • a group for example, the new response graph group may include, but is not limited to, 9 function combinations as shown in (3), (5)-(12).
  • Each combined response map can be generated from the corresponding Gauss-Hermit moment response map according to the formula.
  • GHM 3 GHM 20 +GHM 02 (6)
  • GHM 2 (GHM 30 +GHM 12 ) 2 +(GHM 03 +GHM 21 ) 2 (7)
  • GHM 4 (GHM 30 -3GHM 12 ) 2 +(GHM 03 -3GHM 21 ) 2 (8)
  • GHM 5 (GHM 30 +GHM 12 ) 2 -(GHM 03 +GHM 21 ) 2 (9)
  • the functions shown in the above function combinations (3), (5) to (12) have different degrees of noise resistance, and the combination of computational complexity and image representation ability to achieve noise resistance, computational complexity and image representation ability. balance. As shown in (10)-(12), it can be a function composed of a single low-order Gauss-Hermitian moment, which has strong anti-noise, low computational complexity and low image representation ability.
  • the noise resistance of (5) to (9) increases sequentially, and the computational complexity and image representation ability increase sequentially.
  • the functions including (3), (5)-(12) have appropriate anti-noise, computational complexity and image representation ability.
  • the above-described functional combinations are exemplary, and the selection of functional combinations can be made based on experiments, or can also be made based on experience or other criteria.
  • the function combination it should also be noted that the Gauss-Hermitian moments of different orders have different numerical scale units. For example , the value of GHM 00 is between 0 and tens of thousands, and the value of GHM 10 and GHM 01 may be in the Between thousands and millions, so the Gauss-Hermit moments of different orders need to be added and subtracted after unifying the numerical scale.
  • step S503 a local area is divided in the image response map obtained by using the second set of filter templates based on the determined positions and scales of the key points.
  • step S504 sub-regions are further divided into image regions for each local region, and feature values of each image region are acquired.
  • a local area is obtained to calculate from the area. Extract features.
  • the above-mentioned local area is centered on the key point, and the size is the scale of the key point multiplied by 13.
  • At least one sub-region is divided for each of the above-mentioned local regions, for example, 4 sub-regions are divided in grid form, as shown in FIG. It should be pointed out that this division is only exemplary, and other division manners may also be adopted, for example, 2 equal division, 4 equal division, 9 equal division, etc. are performed individually or simultaneously.
  • the response value of the pixels in each sub-area is calculated to obtain a sub-area value
  • the sub-area value may be the total value of the response value, the mean value or the gradient value, and the like.
  • each template obtains an eigenvalue, so the feature of the pixel will be in the form of a vector, which contains n elements, each element corresponds to the eigenvalue of a template, and then The eigenvalues of a region can also be in the form of a vector with n elements.
  • step S505 the image area feature value of each image area is compared with the specific feature value to set the feature indicator of the image area based on the comparison result; and in step S506, the feature indicator of each image area is combined to determine the image Characteristics.
  • the sub-region values of all sub-regions are binary compared in a certain way to generate a set of multi-dimensional 0 and 1 values as the image features of the key points to which the local region belongs.
  • binary comparison There are also multiple options for the above-mentioned binary comparison. For example, two sub-regions are compared in turn, or one sub-region is selected as a benchmark and compared with other sub-regions.
  • the binary generation of the two sub-regions can also be 1 when the reference sub-region is greater than or equal to the other sub-region, and 0 otherwise, and vice versa.
  • Binary encoding refers to the mutual comparison of feature values between two regions, that is, the corresponding comparison of each element in the feature vector, so the encoding result obtained from one comparison is also a vector, containing n elements, 0 or 1.
  • key point detection for any image and feature extraction based on each key point can be completed as a feature representing the image. And then use the features to perform feature matching between images.
  • a large amount of strong noise in the image may change the actual value of the Gauss-Hermit moment of pixels in a certain area, but has little effect on the magnitude relationship of the Gauss-Hermit moment values between the regions. Therefore, the binary comparison of the local area can greatly suppress the noise information in the image.
  • the image area is divided into sub-areas and compared in order to better obtain the detailed information of the image area, that is, the relationship between each sub-area and other sub-areas is used to express the part of the image area that is different from other image areas, that is characteristics of this image region. In this way, the noise in the image area to which each key point belongs can be suppressed by means of binary comparison, and the difference between each image area and other image areas can be distinguished by the relationship between all sub-areas within it.
  • FIG. 8 shows an image feature extraction process according to the second embodiment of the present application. Among them, the determination of image key points and the extraction of image features are carried out in parallel. Wherein, preferably, the filter template of the Gauss-Hermit moment is used to determine the key points of the image and extract the features of the image area.
  • step S801 the image is convolved with the filter template to determine the image response map.
  • the filter templates include both filter templates for determining image key points and filter templates for image region feature extraction.
  • the filter template group it is preferable to use templates of 10 types of Gauss-Hermitian moments as the filter template group, wherein GHM 00 , GHM 10 , GHM 01 , GHM 30 , GHM 03 , and GHM with fixed sizes 21 , GHM 12 has a total of 7 templates, and its ⁇ is preferably 1.02, and the filter templates GHM 20 , GHM 02 , and GHM 11 with variable size, and its ⁇ is preferably a multiple of 1.02.
  • the order of the filter template can be selected as described above.
  • a response graph group is obtained under the condition that ⁇ of all templates is set to be 1.02, such as the first response graph group in FIG. 9 .
  • step S802 the image is convolved with the first set of templates in the filter templates to determine the position and scale of key points.
  • the position and scale of the key points are determined by using the filter template with variable size in the filter template as the first set of templates, the specific operation of which may be similar to that described above for step S501, which will not be detailed here. describe.
  • a set of response maps is constructed based on at least one response map in the first set of response maps obtained by using the first set of filter templates and the response maps obtained by using the second set of filter templates.
  • a set of response maps is obtained by combining at least one of the response maps obtained in the first set of templates set to a fixed size and the response maps obtained by using the second set of filter templates in step S801, so as to be used based on the key
  • the point information is used to determine the image area, and then the key point features are extracted.
  • the calculation processing unit acquires the position and scale information of the key points in the image
  • the GHM 00 , GHM 10 , GHM 01 , GHM 30 , GHM 03 , GHM 21 , GHM with fixed size 12 and ⁇ are the same as the above GHM 00 , GHM 20 , GHM 02 , GHM 11, a total of 10 template convolutions obtained by the Gaussian-Hermitian moment response graph group 1 (as shown in Figure 3) to perform internal function combination to generate a new
  • the responses shown in Figure 9 are group 2.
  • at least one fixed response graph in the first response graph group can be directly used in the second response graph group, and the rest are new response graphs generated by the new combination.
  • n1 in the second response graph group can be directly the response graph 1 in the first response graph group, and the rest of the response graphs can be obtained by combination, for example, through the above formulas (3), (5) ⁇ (12) obtained by combining at least some of the functions shown.
  • step S804 a local area is divided in the image response map obtained by using the second set of filter templates based on the determined positions and scales of the key points.
  • step S805 sub-regions are further divided into image regions for each local region, and feature values of each image region are acquired.
  • step S806 the image region feature value of each of the at least two image regions is compared with a specific feature value to set a feature indicator of the image region based on the comparison result; and in step S807, the respective image regions are combined Characteristic indicators of the image area to determine the characteristics of the image.
  • the operations of the above steps S804 to S807 may be similar to the operations of the above steps S503 to S506, and will not be described in detail here.
  • the present disclosure first uses appropriate feature extraction means (for example, using Gauss-Hermit moments) to convert the original image information containing a large amount of useless noise information into images with a certain resistance
  • appropriate feature extraction means for example, using Gauss-Hermit moments
  • the basic sub-features of noise and image representation ability are then encoded and combined to generate image features.
  • the noise immunity and image representation ability of this image feature are greatly enhanced.
  • This feature can well express the effective information of high-resolution images, thereby greatly improving the accuracy of image matching and recognition based on this feature.
  • the image feature extraction solution according to the present disclosure can be applied in various application scenarios.
  • video surveillance systems have been used to capture and record video in a large number of public and private locations, such as airports, train stations, supermarkets, homes, and other locations where people, vehicles, etc. are present.
  • surveillance cameras capture places where a large number of objects exist for a long time, and video of the captured objects is recorded into data so that the past presence of specific persons, vehicles, etc. can be retrieved and checked.
  • Such an implementation can be widely used for real-time monitoring, person tracking, vehicle tracking, and the like.
  • the techniques of the present disclosure can be applied to various products.
  • the technology of the present disclosure can be applied to various camera devices, such as lenses mounted on portable devices, photography devices on drones, photography devices in monitoring equipment, and the like.
  • the techniques of the present disclosure can be applied to the camera device itself, eg, built into a camera lens, integrated with the camera lens, so that the techniques of the present disclosure can be in the form of a software program for execution by the camera's processor, or integrated together in the form of an integrated circuit, a processor; or used in a device connected to a camera device, such as a portable mobile device installed with the camera device, so that the technology of the present disclosure can be in the form of a software program to be used by the camera device It can be executed by the processor of the device, or integrated in the form of an integrated circuit, a processor, or even integrated in an existing processing circuit, for example, it can be used for feature extraction during the photographing process.
  • the present invention can be used in many applications.
  • the present invention may be used to monitor, identify, track objects in still images or moving video captured by cameras, and is particularly advantageous for camera-equipped portable devices, (camera-based) mobile phones, and the like.
  • the above-described series of processes and devices may also be implemented by software and/or firmware.
  • the program constituting the software is installed from a storage medium or a network to a computer having a dedicated hardware configuration, such as a general-purpose personal computer 1300 shown in FIG. 10, in which various programs are installed. can perform various functions and so on.
  • 10 is a block diagram showing an example structure of a personal computer of an information processing apparatus that can be employed in an embodiment of the present disclosure.
  • the personal computer may correspond to the above-described exemplary transmitting device or terminal-side electronic device according to the present disclosure.
  • a central processing unit (CPU) 1301 executes various processes according to a program stored in a read only memory (ROM) 1302 or a program loaded from a storage section 1308 to a random access memory (RAM) 1303 .
  • ROM read only memory
  • RAM random access memory
  • data required when the CPU 1301 executes various processes and the like is also stored as needed.
  • the CPU 1301, the ROM 1302, and the RAM 1303 are connected to each other via a bus 1304.
  • Input/output interface 1305 is also connected to bus 1304 .
  • the following components are connected to the input/output interface 1305: an input section 1306, including a keyboard, a mouse, etc.; an output section 1307, including a display such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1308 , including a hard disk, etc.; and a communication section 1309, including a network interface card such as a LAN card, a modem, and the like.
  • the communication section 1309 performs communication processing via a network such as the Internet.
  • a driver 1310 is also connected to the input/output interface 1305 as required.
  • a removable medium 1311 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is mounted on the drive 1310 as needed, so that a computer program read therefrom is installed into the storage section 1308 as needed.
  • a program constituting the software is installed from a network such as the Internet or a storage medium such as a removable medium 1311.
  • a storage medium is not limited to the removable medium 1311 shown in FIG. 10 in which the program is stored and distributed separately from the device to provide the program to the user.
  • the removable medium 1311 include magnetic disks (including floppy disks (registered trademark)), optical disks (including compact disk read only memory (CD-ROM) and digital versatile disks (DVD)), magneto-optical disks (including minidiscs (MD) (registered trademark) )) and semiconductor memory.
  • the storage medium may be the ROM 1302, a hard disk contained in the storage section 1308, or the like, in which programs are stored and distributed to users together with the devices containing them.
  • the methods and systems of the present invention may be implemented in a variety of ways.
  • the methods and systems of the present invention may be implemented by software, hardware, firmware, or any combination thereof.
  • the order of the steps of the method described above is merely illustrative, and unless specifically stated otherwise, the steps of the method of the present invention are not limited to the order specifically described above.
  • the present invention may also be embodied as a program recorded in a recording medium, comprising machine-readable instructions for implementing the method according to the present invention. Accordingly, the invention also covers a recording medium storing a program for implementing the method according to the invention.
  • Such storage media may include, but are not limited to, floppy disks, optical disks, magneto-optical disks, memory cards, memory sticks, and the like.
  • embodiments of the present disclosure may also include the following Illustrative Examples (EE).
  • An image feature extraction device comprising a processing circuit configured to:
  • the feature indicators for each image region are combined to determine a feature of the image.
  • EE 2 The image feature extraction device according to EE1, wherein the feature of the image is a feature in the form of a vector composed of feature indicators of each image area.
  • the image feature extraction device comprises: sequentially comparing the image region feature values of the at least two image regions pairwise.
  • EE5 The image feature extraction apparatus according to EE1, wherein the comparing comprises comparing an image region feature value of each of the at least two image regions with a reference feature value.
  • EE6 The image feature extraction apparatus according to EE5, wherein the reference feature value is a feature value of a selected reference image region among the at least two image regions.
  • the image is divided into the at least two image regions based on the detected position of each keypoint and the keypoint scale information.
  • the image feature extraction apparatus wherein the image region is a local region in the image surrounding the detected keypoint and having a size based on the detected keypoint scale information.
  • the image feature extraction device wherein the image area is a sub-area included in a local area around the detected key point and having a size based on the detected key point scale information in the image .
  • EE11 The graphic feature extraction device according to EE8, wherein keypoint detection is performed based on Gauss-Hermit moments, and wherein the feature of the image region is based on Gauss-Hermit moments of pixels in the image region determined by the characteristics.
  • An image feature extraction method comprising:
  • the feature indicators for each image region are combined to determine a feature of the image.
  • EE13 The image feature extraction method according to EE12, wherein the feature of the image is a feature in the form of a vector composed of feature indicators of each image region.
  • EE14 The image feature extraction method according to EE12, wherein the feature of the image is obtained by concatenating feature indicators of each image region.
  • EE15 The image feature extraction method according to EE12, wherein the comparison comprises: sequentially comparing the image region feature values of the at least two image regions pairwise.
  • EE16 The image feature extraction method according to EE12, wherein the comparing comprises comparing an image region feature value of each of the at least two image regions with a reference feature value.
  • EE17 The image feature extraction method according to EE16, wherein the reference feature value is a feature value of a selected reference image region in the at least two image regions.
  • the image is divided into the at least two image regions based on the detected position of each keypoint and the keypoint scale information.
  • EE20 The image feature extraction method according to EE19, wherein the image region is a local region in the image surrounding the detected keypoint and having a size based on the detected keypoint scale information.
  • EE21 The image feature extraction method according to EE19, wherein the image area is a sub-area included in a local area around the detected key point and having a size based on the detected key point scale information in the image .
  • an image feature extraction device comprising
  • At least one storage device on which instructions are stored which, when executed by the at least one processor, cause the at least one processor to perform the method according to any of EE12-22 .
  • a storage medium storing instructions which, when executed by a processor, enable the method according to EE12-22 to be performed.
  • An image feature extraction apparatus comprising means for performing the method of EE12-22.

Abstract

本公开涉及图像特征提取方法和设备。提出了一种图像特征提取设备，包括处理电路，所述处理电路被配置为：将图像划分为至少两个图像区域；将所述至少两个图像区域中的各图像区域的图像区域特征值与特定特征值进行比较，以基于比较结果来设定图像区域的特征指示符；以及组合各图像区域的特征指示符以确定图像的特征。

Description

图像特征提取方法和设备 技术领域
本公开涉及图像特征提取。
背景技术
近年来，静态图像或一系列运动图像(诸如视频)中的对象检测/识别/比对/跟踪被普遍地和重要地应用于图像处理、计算机视觉和识别领域，并且在其中起到重要作用。对象可以是人的身体部位，诸如脸部、手部、身体等，其它生物或者植物，或者任何其它希望检测的物体。对象识别/验证是最重要的计算机视觉任务之一，其目标是根据输入的照片/视频来准确地识别或验证其中的特定对象。特别地，在拍摄对象图像进行对象识别/验证时，往往是获得图像特征/对象特征，继而基于所获得的图像/对象特征进行对象识别/验证。
然而，当前摄像装置所获得的图像中往往会包含各种噪声，而噪声的存在使得图像质量变差，可能会导致无法提取到足够准确的图像特征/对象特征，继而会影响对象的识别。
因此，需要改进的技术来进行图像特征提取。
除非另有说明，否则不应假定本节中描述的任何方法仅仅因为包含在本节中而成为现有技术。同样，除非另有说明，否则关于一种或多种方法所认识出的问题不应在本节的基础上假定在任何现有技术中都认识到。
发明内容
本公开的一个目的是改进图像特征提取以实现高精度的特征提取，尤其是在存在大量噪声的情况下，继而能够改进对象识别。
特别地，本公开的方案通过提取图像中包含的图像区域的特征并且对所提取的图像区域特征进行编码组合，能够抑制或者消除噪声的影响，提取出更加精确的图像特征。
在一个方面，提供了一种图像特征提取设备，包括处理电路，所述处理电路可以被配置为：将图像划分为至少两个图像区域；将所述至少两个图像区域中的各图像区域的图像区域特征值与特定特征值进行比较，以基于比较结果来设定图像区域的特征 指示符；以及组合各图像区域的特征指示符以确定图像的特征。
在另一个方面，提供了一种图像特征提取方法，该方法可包括将图像划分为至少两个图像区域；将所述至少两个图像区域中的各图像区域的图像区域特征值与特定特征值进行比较，以基于比较结果来设定图像区域的特征指示符；以及组合各图像区域的特征指示符以确定图像的特征。
在还另一方面，提供了一种包括至少一个处理器和至少一个存储设备，所述至少一个存储设备其上存储有指令，该指令在由所述至少一个处理器执行时可使得所述至少一个处理器执行如本文所述的方法。
在仍另一方面，提供了一种存储有指令的存储介质，该指令在由处理器执行时可以使得执行如本文所述的方法。
在仍另一方面，提供了一种包括用于执行如本文所述的方法的部件的装置。
从参照附图的示例性实施例的以下描述，本发明的其它特征将变得清晰。
附图说明
并入说明书中并且构成说明书的一部分的附图示出了本发明的实施例，并且与描述一起用于解释本发明的原理。在附图中，相似的附图标记指示相似的项目。
图1示出了一般性的图像特征提取操作流程。
图2示出了根据本公开的图像特征提取方法的流程图。
图3示出了根据本公开的一个优选实施例的图像特征提取流程。
图4示出了根据本公开的图像特征提取设备的框图。
图5示出了根据本公开的第一实施例的示例性图像特征提取流程。
图6示出了根据本公开的第一实施例的图像特征提取流程的中间输出。
图7示出了根据本公开的实施例的关键点周围的图像区域。
图8示出了根据本公开的第二实施例的图像特征提取流程。
图9示出了根据本公开的第二实施例的图像特征提取流程的中间输出。
图10示出了示出了能够实现本发明的实施例的计算机系统的示例性硬件配置的框图。
虽然在本公开内容中所描述的实施例可能易于有各种修改和另选形式，但是其具体实施例在附图中作为例子示出并且在本文中被详细描述。但是，应当理解，附图以及对其的详细描述不是要将实施例限定到所公开的特定形式，而是相反，目的是要涵 盖属于权利要求的精神和范围内的所有修改、等同和另选方案。
具体实施方式
在下文中将结合附图对本公开的示范性实施例进行描述。为了清楚和简明起见，在说明书中并未描述实施例的所有特征。然而，应了解，在对实施例进行实施的过程中往往需要做出很多特定于实施方式的设置，以便实现开发人员的具体目标，例如，遵从与装置及业务相关的那些限制条件，并且这些限制条件可能会随着实施方式的不同而有所改变。此外，还应该了解，虽然开发工作有可能是非常复杂和费时的，但对得益于本公开内容的本领域技术人员来说，这种开发工作仅仅是例行的任务。
在此，还应当注意，为了避免因不必要的细节而模糊了本公开，在附图中仅仅示出了与至少根据本公开的方案密切相关的处理步骤和/或设备结构，而省略了与本公开关系不大的其他细节。
以下将参照附图来详细描述本发明的实施例。应注意，在附图中相似的附图标记和字母指示相似的项目，并且因此一旦一个项目在一个附图中被定义，则对于随后的附图无需再对其进行论述。
在本公开中，术语“第一”、“第二”等仅仅用于区分元件或者步骤，而不是要指示时间顺序、优先选择或者重要性。
如前文所述的，图像特征提取广泛地应用于各种应用，诸如对象识别、检测、验证、比对、追踪等等，并且起到重要作用。在本文中，对象可以是人的身体部位，诸如脸部、手部、身体等，其它生物或者植物，或者任何其它希望检测的物体。通常，在各种应用操作中需要对图像进行特征提取，继而基于所提取的特征进行各种后续处理/操作。以下将参照图1来描述一般性的图像特征提取操作流程。
首先，进行图像获取来获得包含待识别/验证的对象的图像。图像获取可通过各种已知的图像获取设备来实现，诸如照相机、摄像头、无人机等等，这里将不再详细描述。
在本公开的上下文中，图像可指的是多种图像中的任一种，诸如彩色图像、灰度图像等。应指出，在本说明书的上下文中，图像的类型未被具体限制，只要这样的图像可经受处理以便从中提取图像特征/对象特征、继而可用于识别/检测图像中的对象即可。
此外，图像可以是任何适当的图像，例如由摄像装置获得的原始图像，或者已对 原始图像进行过特定处理/预处理的图像，例如初步过滤，去混叠、畸变消除、对齐、颜色调整、对比度调整、规范化等等。应指出，该特定处理还可以包括本领域已知的其它类型的预处理操作，这里将不再详细描述。应指出，即使不采用上述特定的处理，仍可实现本申请的有利技术效果。
然后，对于所获得的图像来提取图像特征，特别地可以是图像中包含的待识别对象的对象特征，然后输出所提取的对象特征以用于后续操作。
在本说明书的上下文中，图像包含对象指的是图像含有该对象的对象图像。该对象图像有时也可被称为图像中的对象区域。对象识别也即指的是对象区域中的图像进行识别。因此，在本公开的上下文中，特别是在对象识别/验证/跟踪等情况下，需要准确获取对象特征以进行对象识别/验证/跟踪等，在此情况下，图像特征往往指的是待识别/验证对象的对象特征，图像特征提取往往可指的是待识别/验证/跟踪对象的对象特征提取，除非另外指出。
作为示例，在此情况下，可以首先获取图像中的对象区域/对象图像，然后针对所获取的对象区域/对象图像进行特征提取。例如，在拍摄具有对象的场景的图像时，可以首先对拍摄图像进行处理以获得其中的对象的图像，例如对图像进行分割等等，然后主要针对以对象为主体的对象区域/对象图像进行特征提取。
图像特征/对象特征可以采用各种适当方式来表示。作为示例，可以以矢量形式来表示对象的特征，尤其是代表性特征，这可被称为是对象的“特征矢量”。例如在检测脸部的情况下，将选取人脸的代表性部分的像素纹理信息、位置坐标等作为特征来构成图像的特征矢量。由此，基于所获得的特征矢量，可以进行对象识别/检测/跟踪。应指出，特征矢量可根据对象识别中所使用的模型而有所不同，而且并不特别被限制。特别地，图像特征/对象特征还可以涉及图像的其它属性、特性或者以其它方式来表述，而并不局限于此。
由此，基于所提取到的图像特征/对象特征进行各种对象识别/验证等等处理。作为示例，对象识别指的是对于输入的图像，根据提取的图像特征来判断所要识别的图像中的对象是否是某一特定对象；对象认证/验证指的是对于输入的至少两个图像，根据提取的至少两个图像的图像特征之间的相似性来判断各个图像中的对象是否是同一对象。
高质量图像特征/对象特征对于后续处理结果是非常重要的。现在的高解像度图像获取设备获取的图像在捕捉到大量细节的同时常伴随有大量的噪声。图像区域内的 大量噪声信息严重干扰了图像特征的提取，导致现有的特征提取方法无法提取到特征或所提取的特征无法有效地表征图像，从而严重影响后续的基于图像特征的识别、分类等工作。也就是说，现有的特征提取方法通过图像像素无法提取到准确的/高质量的特征，例如无法直接提取到抗噪性强和/或图像表述能力强的图像特征。
特别地，指纹，人脸，人眼虹膜等图像的识别、分类广泛地用于安防领域，需要提供高质量的特征以保证高的分类识别精度，然而指纹，人脸，人眼虹膜等局部纹理信息较少，在其高解像度图像中的任意局部区域内噪声信息明显多过有用的细节纹理信息，现有的特征提取方法如尺度不变特征变换(SIFT)，加速稳健特征(SURF)等都无法从高解像度指纹等图像中成功提取到准确的/高质量的特征，从而会对于分类识别精度造成不利影响。
本公开提出了一种改进的图像特征提取方案。特别地，本公开提出了提取图像中包含的图像区域的特征，并且基于图像区域的图像区域特征值与特定特征值的相对关系来确定图像的特征。特别的，图像中存在大量噪声、尤其是大量的强噪声的情况下，图像噪声可能改变一定局部区域内像素的特征值，但是区域特征值之间的相对关系可适当地减小甚至消除这样的噪声影响，特别地，通过利用相对关系来表征图像特征可以有效地降低、甚至消除噪声对于图像局部的影响，使得所提取的图像特征抗噪性强，能够更加准确地表征图像。
以下将参照附图来详细描述根据本公开的实施例的图像特征提取方法和设备。
图2示出了根据本公开的实施例的图像特征提取设备的流程图。首先，在步骤S201，将图像划分为至少两个图像区域。
根据本公开的实施例，如前所述，图像可以是原始图像或者经过预处理的图像。根据示例，图像可以特别是对象图像，例如，可以是以待识别对象为主体的图像。
根据本公开的实施例，图像可以被适当划分以获得至少两个图像区域。图像区域可以包括至少一个像素，优选地根据图像区域的形状而包含适当数量的像素。根据本公开的实施例，图像区域可以按照任何方式被划分，只要能够覆盖整个图像即可，特别地在彼此基本相邻且不重叠的情况下覆盖整个图像。根据实施例，所划分成的各个图像区域的形状是相同的。例如，图像区域可以为任何适当的形状，例如正方形、正五边形、正六边形、长方形等等。根据实施例，图像区域的确定还可以依赖于图像区域特征的提取方式，下文将详细进行描述。
然后，在步骤S202，将所述至少两个图像区域中的各图像区域的图像区域特征 值与特定特征值进行比较，以基于比较结果来设定图像区域的特征指示符。
根据本公开的实施例，图像区域的特征可以采用各种已知的方式来提取。优选的，图像区域特征可以根据遵从高斯分布的函数的模型来提取，从而获得相对稳健的特征。优选地，图像区域特征可以根据遵从高斯-赫米特矩的模型来提取。
根据本公开的实施例，图像区域特征可以采用各种方式来表示。例如，可以采用矢量的方式，其中矢量中的每个元素对应于一个特征元素的值。作为另一示例，图像区域特征可以由各个特征元素的值的特定组合值，例如统计值，加权统计值等来表示。这样，图像区域特征可以用更精简的值表示，提高存储和传输效率等。根据本公开的实施例，图像区域的特征值可以为该图像区域中所包含的各像素的特征的数学统计值。根据一个示例，图像区域的特征是基于该图像区域中的像素的基于高斯函数的统计特征而确定的。根据另一示例，图像区域的特征是基于该图像区域中的像素的高斯-赫米特矩特征而确定的。
根据实施例，特定特征值可以采用适当的方式来设定。根据实施例，特定特征值可以被预先设定，例如根据先前训练、检测结果被设定，例如为与待识别对象的类别有关的特征值等等。根据实施例，特定特征值可以与图像区域的特征有关。作为一个示例，所述特定特征与该至少两个图像区域中所选定的基准图像区域的特征有关。根据另一示例，所述特定特征值与该至少两个图像区域中的至少一个图像区域的特征有关，诸如其特征或者特征的数学统计值。
根据一个实施例，该特定特征值可以是应用于所有图像区域的，其可以是预先设定的，或者根据所有图像区域的特征值进行统计计算得到的，例如均值，中值等等。根据还另一实施例，特定特征值可以因图像区域而变化，例如对于一个图像区域，特定特征值为其的处于特定方位的至少一个相邻区域的特征值。例如，在网格形式分割的情况下，图像区域的周围八个图像区域中的至少一个可以作为其相邻区域。如果相邻区域为一个相邻区域，则特定特征值可以为该相邻区域的特征值；如果相邻区域多于一个时，则特定特征值可以为该多个相邻区域的特征值的数学统计值，例如均值、中值等等。
由此，每个图像区域的特征值可以与对应于该图像区域的特定特征值进行比较，从而基于比较结果而获得该图像区域的特征。
根据实施例，为了方便进行比较，图像区域特征与特定特征值采用相同的表达方式。作为示例，在图像区域特征以矢量形式表示的情况下，特定特征值可具有相同的 矢量形式，即具有相同数量的矢量元素，包含相同类别的矢量元素等等。从而图像区域特征与特定特征值之间的比较可以对应于矢量中的对应矢量元素进行比较。
根据实施例，比较结果指示图像区域特征与特定特征值之间的相对关系。特征指示符可依赖于特征值/比较结果的表现形式而被以适当的形式表示。作为示例，该特征指示符可以为矢量形式，或者由矢量元素值进行统计而得到的统计值。在矢量形式下，矢量中的每个元素为图像区域特征和特定特征值的相应元素之间的比较结果，该比较结果指示这两者元素之间的相对关系。根据实施例，特征指示符可以采用各种方式来表示，优选地为二进制形式。例如，特征指示符可以用二进制形式表示一个区域的特征是否优于特定特征值，例如，可以用1表示优于，0表示不优于。作为另一示例，特征指示符可以用比值来指示该区域的特征与该特定特征值之间的相对关系，例如，可以指示该区域的特征与该特定特征值之间的比率/比值，例如每个元素之间的比率/比值，或者特征的统计值之间的比率/比值。应指出，比较可对应于一种编码方式，产生相对关系的其它方式也是适用的。
最后，在步骤S203，组合各图像区域的特征指示符以确定图像的特征。
根据实施例，图像的特征为由各个图像区域的特征指示符组成的矢量形式的特征。作为示例，通过将各图像区域的特征指示符进行级联以得到图像的特征，从而可以由图像区域的特征通过编码组合的方式构建得到更高维的图像特征。
优选地，组合方式为基于局部二进制比较的编码组合方式。根据实施例，比较结果、即图像区域的特征指示符为二进制表示，并且通过将各图像区域的特征指示符进行级联以得到图像的特征。根据实施例，图像区域的特征值为矢量形式，而图像区域的特征指示符为相应的二进制矢量形式。这样，图像区域之间进行二进制比较后再进行编码组合可提升抗噪的二进制比较结果的图像表述能力。
特别地，利用特征值之间的相对关系、而不是特征值本身来表征图像特征，类似于对于图像区域特征的编码进行编码组合，这样所获得的图像特征能够利用相对关系来更好地抑制或者消除噪音和多余信息的干扰，抑制甚至消除通常存在于特征值本身中的噪声对于图像特征的影响。
以下将描述根据本公开的一个优选实施例的图像区域确定，其中图像区域的确定还可以依赖于图像特征的提取方式。
根据本公开的实施例，图像区域可以根据图像的关键点分布来进行划分。关键点指的是图像中的指示图像的代表性特征的相应位置。例如，对于人脸图像，可能人的 五官对应的位置以及一些表征人脸姿态的关键位置可被称为关键点。当然，关键点可以指示图像中的其它代表性位置，这里将不再详细描述。根据实施例，通过对图像进行关键点检测，并且基于所检测到的各关键点的位置与关键点尺度信息来将图像划分为所述至少两个图像区域，由此可以根据所确定的关键点来划分图像。这样划分得到的图像区域能够更适当地反映图像的特征分布，继而图像区域特征能够更加适当地表征图像。
以下将参照附图描述根据本公开的示例性图像提取特征的流程图。图3示出了根据本公开的实施例的图像特征提取流程300。
首先，在步骤S301，确定图像的关键点信息，其中包括关键点的位置以及关键点的尺度。
根据本公开的实施例，关键点可通过各种适当的方式而获得。根据实施例，关键点是根据图像的至少一个基础子特征计算取得的。作为示例，所述多个基础子特征可以为基于高斯函数的多个统计特征。作为另一示例，所述多个基础子特征为高斯-赫米特矩特征。
然后，在步骤S302，根据所确定的关键点信息和关键点尺度来确定图像的区域划分。
根据实施例，图像区域可以是图像中的所检测到的关键点周围的且具有基于所检测到的关键点尺度信息的尺寸的局部区域。作为示例，图像区域可以是图像中的环绕所检测到的关键点的区域，关键点可以是图像区域的中心点，该局部区域的尺寸对应于该关键点尺度，例如为关键点尺度的特定倍数。图像区域可以与关键点是对应的。优选地，每个关键点可对应于一个图像区域。特别地，如果有两个或者更多个关键点距离过近，例如小于特定阈值，小于关键点尺度等等，则可以选择这两个或者更多个关键点中的任一个关键点作为基础来构建图像区域。
根据实施例，图像区域可以是图像中所检测到的关键点周围的且具有基于所检测到的关键点尺度信息的尺寸的局部区域所包含的子区域。特别地，图像区域是通过对基于关键点确定的图像局部区域进行进一步划分而得到的子区域。在此情况下，子区域的划分可以采用任何适当的方式来进行，只要其能够适当覆盖局部区域即可。例如，子区域划分可以是简单的网格形式等等，这里将不再详细描述。子区域的特征可通过任何适当的方式来确定，例如上文所述的方式，这里将不再详细描述。
然后，在步骤S303，可以获取划分得到的图像区域的特征。图像区域的特征可 采用各种适当的方式来获取。作为示例，采用遵从高斯分布的模型来进行特征提取，这样基于高斯分布的统计特征兼具一定程度的抗噪能力与图像表述能力。
根据实施例，图像区域的特征可以依赖于关键点的确定而被确定。优选地，图像区域的特征可以采用与关键点确定方式相同的方式来获得。例如，采用与用于获得关键点的模板相同或者相类似的模板来获得。根据一个实施例，基于高斯函数来进行关键点检测，并且其中所述图像区域的特征是基于该图像区域中的像素的基于高斯函数的统计特征而确定的。根据另一个实施例，基于高斯-赫米特矩来进行关键点检测，并且其中所述图像区域的特征是基于该图像区域中的像素的高斯-赫米特矩特征而确定的。应指出，上述图像区域的特征的获取可适用于图像区域对应于局部区域以及局部区域的子区域等等情况。
在步骤S304，将所述至少两个图像区域中的各图像区域的图像区域特征值与特定特征值进行比较，以基于比较结果来设定图像区域的特征指示符；以及在步骤S305，组合各图像区域的特征指示符以确定图像的特征。
在比较过程中，对于所有图像区域可以采用相同的比较规则，例如所有图像区域的特征与同样特定特征值进行比较，或者各个图像区域的特征均与同样相对方位处的相邻区域的特征进行比较。根据示例，比较过程中可以将图像区域进行简单的配对，然后每两个图像区域进行比较。例如，可以将在特定方向上彼此相邻的两个图像区域配对然后进行比较。
从而获得比较结果作为图像区域的特征指示符。特别地，特征指示符可以是二进制表示，或者任何其他适当的相对比值表示，这里将不再详细描述。
根据实施例，在图像区域对应于局部区域的子区域的情况下，可以通过对子区域的特征进行比较来构建子区域特征指示符，从而可获得包含该子区域的局部区域的特征，继而可基于局部区域的特征而获得图像特征。该子区域特征比较也可采用各种适当方式，特别地如上文所述的二进制或比率的方式，来进行，这里将不再详细描述。进一步进行子区域划分并进行比较能够更好地获取局部图像区域的细节信息，即用每个子区域与其他子区域之间的关系来表述该局部图像区域区别于其他局部图像区域的部分，这样，各个关键点所属的局部图像区域内的噪声也可以利用相对关系来抑制或消除，而各局部图像区域与其他局部图像区域的不同可通过其内部的所有子区域之间的编码关系来区分。
最后，将所有图像区域的特征指示符进行组合来得到整个图像特征的表示。作为 示例，整个图像特征可以为矢量形式，其中每个元素对应于一个图像区域的特征指示符。特别地，如果一个图像区域的特征为矢量形式，则该图像区域的特征指示符可能也是矢量形式，继而得到的整个图像特征可以类似于矩阵形式，其中一行或者一列对应于一个图像区域的特征指示符的矢量。
有利效果
在本公开的实施例中，通过以相对值的形式获取图像中的图像区域的特征并且进行适当的组合，可以获得鲁棒的图像特征/对象特征，其中相对值形式能够更好地消除图像中噪声的影响。特别地，采用二进制比较来进行编码组合，二进制比较可大幅度减少噪音和多余信息的干扰。局部二进制比较后编码组合可提升抗噪的二进制比较结果的图像表述能力。
进一步地，通过适当的方式进行图像关键点确定以及特征提取，尤其是基于高斯-赫米特矩进行关键点的确定以及特征提取，从而获得具有一定抗噪性和图像表述能力的基础子特征来进行特征比较，使得进一步提升最终获得的图像特征的抗噪性和图像表述能力。
以下将描述根据本公开的实施例的能够进行图像特征提取的电子设备。图4示出了根据本公开的实施例的用于图像特征提取的电子设备的框图。电子设备400包括处理电路420，该处理电路420可被配置为将图像划分为至少两个图像区域；将所述至少两个图像区域中的各图像区域的图像区域特征值与特定特征值进行比较，以基于比较结果来设定图像区域的特征指示符；以及组合各图像区域的特征指示符以确定图像的特征。
在上述装置的结构示例中，处理电路420可以是通用处理器的形式，也可以是专用处理电路，例如ASIC。例如，处理电路420能够由电路(硬件)或中央处理设备(诸如，中央处理单元(CPU))构造。此外，处理电路420上可以承载用于使电路(硬件)或中央处理设备工作的程序(软件)。该程序能够存储在存储器(诸如，布置在存储器中)或从外面连接的外部存储介质中，以及经由网络(诸如，互联网)下载。
根据本公开的实施例，处理电路420可以包括用于实现上述功能的各个单元，特别地，处理电路420可包括被配置用于将图像划分为至少两个图像区域的划分单元422；被配置用于将所述至少两个图像区域中的各图像区域的图像区域特征值与特定特征值进行比较，以基于比较结果来设定图像区域的特征指示符的比较单元424；以及被 配置用于组合各图像区域的特征指示符以确定图像的特征的组合单元426。进一步地，处理电路420还可包括被配置为确定图像的关键点信息(包括关键点的位置以及关键点的尺度)的关键点确定单元428和被配置为提取图像区域特征的特征提取单元430。每个单元可以进行如上文所述地操作，这里将不再详细描述。
此外，处理电路420还可包括存储单元432，以用于存储由处理电路420产生的各种信息以及其他进行处理操作所需的信息。该存储单元可以是易失性存储器和/或非易失性存储器。例如，存储器可以包括但不限于随机存储存储器(RAM)、动态随机存储存储器(DRAM)、静态随机存取存储器(SRAM)、只读存储器(ROM)、闪存存储器。
附图中的部分单元用虚线绘出，旨在说明该单元并不一定被包含在处理电路中，甚至并不一定是必需的。作为示例，该关键点确定单元428以及特征提取单元430可以在终端侧电子设备中而处理电路之外，甚至可以位于电子设备400之外。特别地，即使不存在关键点确定单元428和特征提取单元430，本申请的电子设备仍可有效地工作并获得上述的有利技术效果。存储单元还可以位于电子设备内但在处理电路之外，或者甚至位于电子设备之外。需要注意的是，尽管图4中将各个单元示为分立的单元，但是这些单元中的一个或多个也可以合并为一个单元，或者拆分为多个单元。
应注意，上述各个单元仅是根据其所实现的具体功能划分的逻辑模块，而不是用于限制具体的实现方式，例如可以以软件、硬件或者软硬件结合的方式来实现。在实际实现时，上述各个单元可被实现为独立的物理实体，或者也可由单个实体(例如，处理器(CPU或DSP等)、集成电路等)来实现。此外，上述各个单元在附图中用虚线示出指示这些单元可以并不实际存在，而它们所实现的操作/功能可由处理电路本身来实现。
应理解，图4仅仅是用于图像特征提取的电子设备的概略性结构配置，电子设备400还可以包括其他可能的部件。可选地，电子设备400还可以包括未示出的其它部件，诸如网络接口、控制器等。处理电路可以与外部存储器相关联。例如，处理电路可以直接或间接(例如，中间可能连接有其它部件)连接到外部存储器，以进行数据的存取。
以下将参照附图详细描述根据本公开的实施例的示例性图像特征提取的操作流程图。其中，图像关键点的确定以及图像特征的提取可以采用不同的方式来执行。
图5示出了根据本申请的第一实施例的图像特征提取流程。其中，图像关键点的 确定以及图像特征的提取是顺序进行的。其中，优选地，基于高斯-赫米特矩进行图像的关键点确定以及图像区域特征的提取。
在步骤S501，利用第一组滤波器模板对图像进行卷积以确定图像响应图中的关键点位置和尺度。应指出，第一组滤波器模板主要是指代用于确定关键点信息的图像处理函数。
作为示例，第一组滤波器模板由特定数量的滤波器模板构成，例如由特定数量的、相同类型但具有不同参数(例如，尺寸、阶数等)的滤波器模板构成。应指出，模板的类型和数量可以被适当地设定，例如可以在考虑了计算时间、计算量、计算复杂性、精度、可识别性等等的至少一种的情况下被设定，以便能够以高成本效率来实现关键点确定。
特别地，第一组滤波器模板包含特定数量的可变尺寸的滤波器模板。根据本公开，在滤波器模板的尺寸设定为特定尺寸的情况下利用第一组滤波器模板来获得模板响应图，以获得关键点位置。其中，所有滤波器模板的尺寸是相同的。根据本公开，特定尺寸可以是根据经验设定的，例如可以是与所要识别的对象的类别有关的经验值，或者可以在考虑了计算精度、可识别性等中的至少一种的情况下被设定，例如如前所述。特别地，模板响应指的是对于每个像素点的响应，例如对于每个像素点会获得一个特征值，然后通过将像素点的特征值之间进行比较、判定，来选择关键点。然后，通过迭代地改变滤波器模板的尺寸来确定关键点的尺度。其中，在由多个可变尺寸的滤波器模板得到的响应图组成的图像尺度空间中通过寻找极值点来确定关键点的尺度。特别地，迭代次数以及每次迭代中尺寸改变步长可根据经验设定，或者在考虑了计算精度、可识别性等中的至少一种的情况下被设定，例如如前所述。
优选地，作为示例，利用高斯-赫米特矩的模板来进行关键点信息的确定。高斯-赫米特矩的定义如下(1)和(2)所示。
Figure PCTCN2021102237-appb-000001
其中
Figure PCTCN2021102237-appb-000002
其中GHM pq为使用宽为2w+1，高为2h+1的p+q阶高斯-赫米特矩滤波器模板从宽为2w+1，高为2h+1的图像区域内计算到的高斯-赫米特矩的响应值，f(x,y)为在图像的x,y位置上的图像像素灰度值，H p(x；σ)，H q(y；σ)为如(2)所示的赫米特函数，σ指示高斯-赫米特矩滤波器模板的大小/尺寸，p，q指示赫米特函数的阶数。由于高斯-赫米特矩兼具高 斯函数的抑制噪声的能力以及赫米特函数的图像表述能力。在本公开中可作为理想的基础子特征用来取代易受噪声影响且图像表述能力不够的像素灰度。
在本公开的上下文中，高斯-赫米特矩的阶数可被适当地设定，例如根据经验设定，或者特别地是考虑了计算量、计算复杂性、可识别性等性能指标之间的折中而被设定。作为示例，考虑到计算复杂度、抗噪性以及图像表述能力的平衡，在本方法中优选为使用阶数从0到3阶的高斯-赫米特矩。更高阶数如超过4阶的高斯-赫米特矩尽管具有更高的图像表述能力，但其计算复杂度明显升高且抑制噪声的能力有所减弱。
根据本公开，采用具有可变尺寸的至少一个高斯-赫米特矩的滤波器模板作为第一组滤波器模板来进行图像关键点检测。作为示例，该第一组滤波器模板可包括可变尺寸的GHM 20，GHM 02，GHM 11模板。其σ优选为1.02的倍数。应指出，其它可变尺寸的模板也是可能的。
其中由多个可变尺寸的GHM 20，GHM 02，GHM 11模板卷积得到的高斯-赫米特矩的响应图用来进行图像关键点检测，如图6中的响应图1，响应图2和响应图3所示。尽管可以采用任意阶数的高斯-赫米特矩或高斯-赫米特矩的组合来进行关键点检测。但在本实施例优选为使用相同尺寸模板的GHM 20，GHM 02，GHM 11组成的函数作为关键点检测函数。
Figure PCTCN2021102237-appb-000003
Figure PCTCN2021102237-appb-000004
GHM 1＝GHM 10 2+GHM 01 2          (5)
通过改变模板的尺寸，在多个由多个可变尺寸的GHM 20，GHM 02，GHM 11模板卷积得到的高斯-赫米特矩的响应图组成的图像尺度空间内优选为通过寻找(3)所示函数的极值点来确定关键点的尺度，以使特征提取过程对于图像尺度变化具有很高的鲁棒性。应指出，此处的响应图的数量以及模板的数量仅仅是示例性的，并且其它数量的模板和响应图也是可行的，并且通常响应图的数量对应于所使用的模板的数量。
(3)所示函数具有对图像旋转不变的能力以及相当程度的抗噪性以及图像表述能力，可以从图像中检测出足够的独特的关键点。除此之外，其他的高斯-赫米特矩组合的函数如(4)或(5)所示函数也可用来检测关键点，但考虑到受图像旋转变化的鲁棒性或从图像中可检测出的关键点数以及计算复杂性的影响，优选为使用(3)所示函数。
应指出，上述先确定关键点位置然后确定关键点尺度的操作仅是示例性的，优选地，也可在利用可变尺寸的滤波器模板获得了响应图空间之后通过确定函数极值来同时确定关键点信息，特别是关键点位置和关键点尺度。此外，其他获取关键点或者关键点尺度的 方法也可应用到本公开的实施例中，这里将不再详细描述。
在步骤S502，利用第二组滤波器模板对图像进行卷积以获得图像响应图。在此情况下，第二组滤波器模板中的各滤波器模板是固定尺寸的模板。应指出，第二组滤波器模板可以与第一组滤波器模板相同、部分相同、或者完全不同。优选地，图像区域的特征同样是根据与确定关键点的方式相同的方式来被获取。
特别的，对于输入的图像，设定至少一个不同尺寸不同阶的高斯-赫米特矩的滤波器模板并对图像进行卷积滤波，从而获得原始图像对于至少一个不同尺寸不同阶的高斯-赫米特矩的响应图。作为另一示例，第二组滤波器模板中的各滤波器模板的尺寸可相同，但阶数不同。
作为示例，尽管高斯-赫米特矩的阶数与滤波器模板尺寸可以有任意种组合，在本实施例中优选为使用10种类型的高斯-赫米特矩的模板作为第二组滤波器模板，其中具有固定尺寸的GHM 00，GHM 10，GHM 01，GHM 30，GHM 03，GHM 21，GHM 12共7个模板，以及前述第一组滤波器模板GHM 20，GHM 02，GHM 11模板。滤波器模板的阶数可如上所述地设定，在本方法中优选为使用阶数从0到3阶的高斯-赫米特矩。滤波器模板的尺寸可以相同，σ优选为1.02。从而，可以由这10个模板获得图像的响应图，如图6中的响应图n1，n2，n3，…，n9所示。应指出，响应图n1，n2，n3，…，n9的数量仅是示例，其还可包含任何其他数量的响应图，该数量可以被预先设定，或者在考虑了抗噪性，计算复杂度和图像表述能力的平衡的情况下被设定。
优选地，根据本公开的实施例，首先将具有固定尺寸的GHM 00，GHM 10，GHM 01，GHM 30，GHM 03，GHM 21，GHM 12以及σ与上述GHM 00等相同的GHM 20，GHM 02，GHM 11共10个模板卷积图像从而分别获得高斯-赫米特矩的响应图组，其中包括与各个模板一一对应的10个响应图，然后将通过进行内部函数组合生成新的响应图组，例如该新的响应图组可包括但不局限于如(3)，(5)～(12)所示的9个函数组合。
每个组合响应图可由相应的高斯-赫米特矩的响应图按照公式生成。
GHM 3＝GHM 20+GHM 02                             (6)
GHM 2＝(GHM 30+GHM 12) 2+(GHM 03+GHM 21) 2             (7)
GHM 4＝(GHM 30-3GHM 12) 2+(GHM 03-3GHM 21) 2           (8)
GHM 5＝(GHM 30+GHM 12) 2-(GHM 03+GHM 21) 2             (9)
GHM 6＝GHM 00                                  (10)
GHM 7＝GHM 10                                  (11)
GHM 8＝GHM 01           (12)
上述函数组合(3)，(5)～(12)所示的函数，具有不同程度的抗噪性，计算复杂度和图像表述能力的组合以实现抗噪性，计算复杂度和图像表述能力的平衡。如(10)～(12)所示可以为由单个低阶高斯-赫米特矩组成的函数，其抗噪性较强，计算复杂度和图像表述能力较低。(5)～(9)的抗噪性依次增强，计算复杂度和图像表述能力依次增高。最终包含(3)，(5)～(12)的函数整体具有适当的抗噪性，计算复杂度及图像表述能力。应指出，上述函数组合是示例性的，函数组合的选择可以根据实验来作出，或者也可以根据经验或其他准则来作出。此外，在选择函数组合时还应注意不同阶的高斯-赫米特矩因具有不同的数值尺度单位，如GHM 00的数值在0到数万之间，GHM 10，GHM 01的数值可能在数千到数百万之间，所以不同阶的高斯-赫米特矩需要在统一数值尺度后进行加减计算。
在步骤S503中，基于所确定的关键点位置和尺度在利用第二组滤波器模板获得的图像响应图中划分出局部区域。在步骤S504中，对于各局部区域进一步划分出子区域以作为图像区域，并获取各图像区域的特征值。
在包含响应图n1，n2，n3，…，n9的响应图组中的每个响应图中根据检测到的每个关键点的位置与尺度信息，都获取一个局部区域用来从该区域中计算提取特征。在本实施例中优选为上述局部区域以关键点为中心，尺寸为关键点尺度乘以13。对于上述每个局部区域都划分出至少一个子区域，例如以网格形式划分出4个子区域，如图7所示，以关键点400为中心的局部区域被划分出子区域401到404。应指出，这种划分仅是示例性的，其它划分方式也是可以采用的，例如单独或同时进行2等分，4等分，9等分等。
然后，对每个子区域内的像素的响应值进行计算获取一个子区域值，该子区域值可以为响应值的总值，均值或梯度值等。理论上，对于一个像素点，使用n个模板，每个模板获得一个特征值，从而该像素点的特征会是矢量形式，其中包含n个元素，每个元素对应于一个模板的特征值，继而区域的特征值也可以是包含n个元素的矢量形式。
在步骤S505，将各图像区域的图像区域特征值与特定特征值进行比较，以基于比较结果来设定图像区域的特征指示符；以及在步骤S506，组合各图像区域的特征指示符以确定图像的特征。
其中，对所有子区域的子区域值以一定方式进行二进制比较生成一组多维的0，1值作为该局部区域所属关键点的图像特征。上述二进制比较方式也可以有多种选择。例如依次两两子区域比较或者选定一个子区域为基准并与其余子区域进行比较等。两个子区域的二 进制生成也可以为基准子区域大于或等于另一子区域为1，否则为0，反之也可。二进制编码指的是两个区域之间的特征值的相互比较，也就是特征矢量中的每个元素的对应比较，从而一次比较得到的编码结果也是一个矢量，包含n个元素，0或者1。通过上述方式即可完成对任意图像的关键点检测以及基于每个关键点的特征提取作为代表该图像的特征。并在之后利用特征进行图像间的特征匹配。
在本实施例中，图像中大量的强噪声可能改变一定区域内像素的高斯-赫米特矩实际数值，但对于区域之间的高斯-赫米特矩值的大小关系影响甚微。因此对局部区域进行二进制比较可大幅度的抑制图像内的噪声信息。同时对图像区域进行子区域划分并进行比较是为了更好的获取图像区域的细节信息，即用每个子区域与其他子区域之间的关系来表述该图像区域区别于其他图像区域的部分，即该图像区域的特征。这样，各个关键点所属图像区域内的噪声可有二进制比较的方式抑制，而各图像区域与其他图像区域的不同可通过其内部的所有子区域之间的关系来区分。
图8示出了根据本申请的第二实施例的图像特征提取流程。其中，图像关键点的确定以及图像特征的提取是并行进行的。其中，优选地，利用高斯-赫米特矩的滤波器模板来进行图像的关键点确定以及图像区域特征的提取。
在步骤S801，利用滤波器模板对图像进行卷积以确定图像响应图。其中，滤波器模板包括用于确定图像关键点的滤波器模板以及用于图像区域特征提取的滤波器模板两者。
作为示例，在本实施例中优选为使用10种类型的高斯-赫米特矩的模板作为滤波器模板组，其中具有固定尺寸的GHM 00，GHM 10，GHM 01，GHM 30，GHM 03，GHM 21，GHM 12共7个模板，其σ优选为1.02，以及具有可变尺寸的滤波器模板GHM 20，GHM 02，GHM 11模板，其σ优选为1.02的倍数。滤波器模板的阶数可如上所述地被选定。由此，在设定所有模板的σ为1.02的情况下获得响应图组，如图9中的第一响应图组。
在步骤S802中，利用滤波器模板中的第一组模板对图像进行卷积以确定关键点位置和尺度。特别地，利用滤波器模板中的具有可变尺寸的滤波器模板作为第一组模板来确定关键点位置和尺度，其具体操作可类似于上文针对步骤S501所述的，这里将不再详细描述。
在步骤S803中，基于第一响应图组中的利用第一组滤波器模板获得的响应图中的至少一个响应图以及利用第二组滤波器模板获得的响应图而构建响应图组。其中，利用在被设定为固定尺寸的第一组模板获取的响应图中的至少一个以及在步骤S801中利用第二组 滤波器模板获得的响应图组合得到响应图组，以用于基于关键点信息来确定图像区域，继而提取关键点特征。
计算处理单元在获取到图像内的关键点的位置与尺度信息后，在本实施例中优选为首先将具有固定尺寸的GHM 00，GHM 10，GHM 01，GHM 30，GHM 03，GHM 21，GHM 12以及σ与上述GHM 00等相同的GHM 20，GHM 02，GHM 11共10个模板卷积获得的高斯-赫米特矩的响应图组1(如图3所示)进行内部函数组合生成新的如图9所示响应图组2。其中，第一响应图组中固定至少一个响应图可以在第二响应图组中直接使用，其余为新组合生成的新响应图。例如第二响应图组中的n1可以是直接是第一响应图组中的响应图1，而其余响应图可以是组合得到的，例如通过上述式子(3)，(5)～(12)所示的函数组合中的至少一些而得到的。
在步骤S804中，基于所确定的关键点位置和尺度在利用第二组滤波器模板获得的图像响应图中划分出局部区域。在步骤S805中，对于各局部区域进一步划分出子区域以作为图像区域，并获取各图像区域的特征值。在步骤S806，将所述至少两个图像区域中的各图像区域的图像区域特征值与特定特征值进行比较，以基于比较结果来设定图像区域的特征指示符；以及在步骤S807，组合各图像区域的特征指示符以确定图像的特征。上述步骤S804到S807的操作可类似于上文步骤S503到S506的操作，这里将不再详细描述。
特别地，本公开在对高解像度图像进行特征提取的过程中，首先利用适当的特征提取手段(例如，利用高斯-赫米特矩)将包含大量无用噪声信息的图像原始信息转换成具有一定抗噪性和图像表述能力的基础子特征，之后对基础子特征进行编码组合来生成图像特征。该图像特征的抗噪性和图像表述能力得到了大幅强化。该特征可很好地表述高解像度图像的有效信息从而大幅提高基于该特征的图像匹配、识别等的精度。从而实现了对于高解像度图像的特征提取与匹配的精度提高，对于指纹，人脸等纹理简单的高解像度图像的特征提取与匹配的精度明显提高，以及可完成基于图像匹配的指纹，人脸等纹理简单的高解像度图像的拼接与识别等工作。
根据本公开的图像特征提取方案可应用于各种应用场景中。在一种常用应用中，视频监控系统已经被用于捕获和记录大量公共和私人场所(诸如机场、火车站、超市、家庭以及有人、车辆等存在的其它场所)的视频。通常，监控照相机捕获长期存在大量对象的场所，并且将所捕获的对象的视频记录到数据中，从而特定人物或车辆等的以往的存在可被检索并且被进行检查。这样的实现可被广泛地用于实时监控、人物跟踪、车辆跟踪等。
本公开的技术能够应用于各种产品。本公开的技术可以应用于各种摄像装置中，例如安装在便携式设备的镜头，无人机上的拍摄装置，监控设备等中的拍摄装置，等等。例如，本公开的技术能够应用于摄像装置本身，例如内置于相机镜头中，与相机镜头集成在一起，这样，本公开的技术可以以软件程序的形式以便由摄像装置的处理器来执行，或者以集成电路、处理器的形式集成在一起；或者用于与摄像装置相连接的设备中，例如安装有该摄像装置的便携式移动设备，这样，本公开的技术可以以软件程序的形式以便由摄像装置的处理器来执行，或者以集成电路、处理器的形式集成在一起，甚至集成在已有的处理电路中，例如可用于在拍照过程中进行特征提取。
本发明可被用于许多应用。例如，本发明可被用于监测、识别、跟踪照相机捕获的静态图像或移动视频中的对象，并且对于配备有相机的便携式设备、(基于相机)的移动电话等等是尤其有利的。
另外，应当理解，上述系列处理和设备也可以通过软件和/或固件实现。在通过软件和/或固件实现的情况下，从存储介质或网络向具有专用硬件结构的计算机，例如图10所示的通用个人计算机1300安装构成该软件的程序，该计算机在安装有各种程序时，能够执行各种功能等等。图10是示出根据本公开的实施例的中可采用的信息处理设备的个人计算机的示例结构的框图。在一个例子中，该个人计算机可以对应于根据本公开的上述示例性发射设备或终端侧电子设备。
在图10中，中央处理单元(CPU)1301根据只读存储器(ROM)1302中存储的程序或从存储部分1308加载到随机存取存储器(RAM)1303的程序执行各种处理。在RAM 1303中，也根据需要存储当CPU 1301执行各种处理等时所需的数据。
CPU 1301、ROM 1302和RAM 1303经由总线1304彼此连接。输入/输出接口1305也连接到总线1304。
下述部件连接到输入/输出接口1305：输入部分1306，包括键盘、鼠标等；输出部分1307，包括显示器，比如阴极射线管(CRT)、液晶显示器(LCD)等，和扬声器等；存储部分1308，包括硬盘等；和通信部分1309，包括网络接口卡比如LAN卡、调制解调器等。通信部分1309经由网络比如因特网执行通信处理。
根据需要，驱动器1310也连接到输入/输出接口1305。可拆卸介质1311比如磁盘、光盘、磁光盘、半导体存储器等等根据需要被安装在驱动器1310上，使得从中读出的计算机程序根据需要被安装到存储部分1308中。
在通过软件实现上述系列处理的情况下，从网络比如因特网或存储介质比如可拆 卸介质1311安装构成软件的程序。
本领域技术人员应当理解，这种存储介质不局限于图10所示的其中存储有程序、与设备相分离地分发以向用户提供程序的可拆卸介质1311。可拆卸介质1311的例子包含磁盘(包含软盘(注册商标))、光盘(包含光盘只读存储器(CD-ROM)和数字通用盘(DVD))、磁光盘(包含迷你盘(MD)(注册商标))和半导体存储器。或者，存储介质可以是ROM 1302、存储部分1308中包含的硬盘等等，其中存有程序，并且与包含它们的设备一起被分发给用户。
应指出，文中所述的方法和设备可被实现为软件、固件、硬件或它们的任何组合。有些组件可例如被实现为在数字信号处理器或者微处理器上运行的软件。其他组件可例如实现为硬件和/或专用集成电路。
另外，可采用多种方式来实行本发明的方法和系统。例如，可通过软件、硬件、固件或它们的任何组合来实行本发明的方法和系统。上文所述的该方法的步骤的顺序仅是说明性的，并且除非另外具体说明，否则本发明的方法的步骤不限于上文具体描述的顺序。此外，在一些实施例中，本发明还可具体化为记录介质中记录的程序，包括用于实施根据本发明的方法的机器可读指令。因此，本发明还涵盖了存储用于实施根据本发明的方法的程序的记录介质。这样的存储介质可以包括但不限于软盘、光盘、磁光盘、存储卡、存储棒等等。
本领域技术人员应当意识到，在上述操作之间的边界仅仅是说明性的。多个操作可以结合成单个操作，单个操作可以分布于附加的操作中，并且操作可以在时间上至少部分重叠地执行。而且，另选的实施例可以包括特定操作的多个实例，并且在其他各种实施例中可以改变操作顺序。但是，其它的修改、变化和替换同样是可能的。因此，本说明书和附图应当被看作是说明性的，而非限制性的。
另外，本公开的实施方式还可以包括以下示意性示例(EE)。
EE 1.一种图像特征提取设备，包括处理电路，所述处理电路被配置为：
将图像划分为至少两个图像区域；
将所述至少两个图像区域中的各图像区域的图像区域特征值与特定特征值进行比较，以基于比较结果来设定图像区域的特征指示符；以及
组合各图像区域的特征指示符以确定图像的特征。
EE 2、根据EE1所述的图像特征提取设备，其中，该图像的特征为由各图像区域的特征指示符组成的矢量形式的特征。
EE3、根据EE1所述的图像特征提取设备，其中，通过将各图像区域的特征指示符进行级联以得到图像的特征。
EE4、根据EE1所述的图像特征提取设备，其中，所述比较包括：将所述至少两个图像区域的图像区域特征值依次两两进行比较。
EE5、根据EE1所述的图像特征提取设备，其中，所述比较包括将所述至少两个图像区域中的各图像区域的图像区域特征值与基准特征值进行比较。
EE6、根据EE5所述的图像特征提取设备，其中，所述基准特征值为所述至少两个图像区域中所选定的基准图像区域的特征值。
EE7、根据EE1所述的图像特征提取设备，其中，图像区域的特征值为矢量形式，而图像区域的特征指示符为相应的二进制矢量形式。
EE8、根据EE1所述的图像特征提取设备，其中，所述处理电路进一步配置为：
对图像进行关键点检测，并且
基于所检测到的各关键点的位置与关键点尺度信息来将图像划分为所述至少两个图像区域。
EE9、根据EE8所述的图像特征提取设备，其中，图像区域是图像中的在所检测到的关键点周围的且具有基于所检测到的关键点尺度信息的尺寸的局部区域。
EE10、根据EE8所述的图像特征提取设备，其中，图像区域是图像中的在所检测到的关键点周围的且具有基于所检测到的关键点尺度信息的尺寸的局部区域所包含的子区域。
EE11、根据EE8所述的图形特征提取设备，其中，基于高斯-赫米特矩来进行关键点检测，并且其中所述图像区域的特征是基于该图像区域中的像素的高斯-赫米特矩特征而确定的。
EE12、一种图像特征提取方法，包括：
将图像划分为至少两个图像区域；
将所述至少两个图像区域中的各图像区域的图像区域特征值与特定特征值进行比较，以基于比较结果来设定图像区域的特征指示符；以及
组合各图像区域的特征指示符以确定图像的特征。
EE13、根据EE12所述的图像特征提取方法，其中，该图像的特征为由各图像区域的特征指示符组成的矢量形式的特征。
EE14、根据EE12所述的图像特征提取方法，其中，通过将各图像区域的特征指 示符进行级联以得到图像的特征。
EE15、根据EE12所述的图像特征提取方法，其中，所述比较包括：将所述至少两个图像区域的图像区域特征值依次两两进行比较。
EE16、根据EE12所述的图像特征提取方法，其中，所述比较包括将所述至少两个图像区域中的各图像区域的图像区域特征值与基准特征值进行比较。
EE17、根据EE16所述的图像特征提取方法，其中，所述基准特征值为所述至少两个图像区域中所选定的基准图像区域的特征值。
EE18、根据EE12所述的图像特征提取方法，其中，图像区域的特征值为矢量形式，而图像区域的特征指示符为相应的二进制矢量形式。
EE19、根据EE12所述的图像特征提取方法，进一步包括：
对图像进行关键点检测，并且
基于所检测到的各关键点的位置与关键点尺度信息来将图像划分为所述至少两个图像区域。
EE20、根据EE19所述的图像特征提取方法，其中，图像区域是图像中的在所检测到的关键点周围的且具有基于所检测到的关键点尺度信息的尺寸的局部区域。
EE21、根据EE19所述的图像特征提取方法，其中，图像区域是图像中的在所检测到的关键点周围的且具有基于所检测到的关键点尺度信息的尺寸的局部区域所包含的子区域。
EE22、根据EE19所述的图形特征提取方法，其中，基于高斯-赫米特矩来进行关键点检测，并且其中所述图像区域的特征是基于该图像区域中的像素的高斯-赫米特矩特征而确定的。
EE23、一种图像特征提取设备，包括
至少一个处理器；和
至少一个存储设备，所述至少一个存储设备在其上存储指令，该指令在由所述至少一个处理器执行时，使所述至少一个处理器执行根据EE12-22中任一项所述的方法。
EE24、一种存储指令的存储介质，该指令在由处理器执行时能使得执行根据EE12-22所述的方法。
EE25、一种包括用于执行EE12-22所述的方法的部件的图像特征提取装置。
虽然已经详细说明了本公开及其优点，但是应当理解在不脱离由所附的权利要求所限定的本公开的精神和范围的情况下可以进行各种改变、替代和变换。而且，本公 开实施例的术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、物品或者设备中还存在另外的相同要素。
虽然已详细描述了本公开的一些具体实施例，但是本领域技术人员应当理解，上述实施例仅是说明性的而不限制本公开的范围。本领域技术人员应该理解，上述实施例可以被组合、修改或替换而不脱离本公开的范围和实质。本公开的范围是通过所附的权利要求限定的

Claims (25)

  1. 一种图像特征提取设备，包括处理电路，所述处理电路被配置为：
    将图像划分为至少两个图像区域；
    将所述至少两个图像区域中的各图像区域的图像区域特征值与特定特征值进行比较，以基于比较结果来设定图像区域的特征指示符；以及
    组合各图像区域的特征指示符以确定图像的特征。
  2. 根据权利要求1所述的图像特征提取设备，其中，该图像的特征为由各图像区域的特征指示符组成的矢量形式的特征。
  3. 根据权利要求1所述的图像特征提取设备，其中，通过将各图像区域的特征指示符进行级联以得到图像的特征。
  4. 根据权利要求1所述的图像特征提取设备，其中，所述比较包括：将所述至少两个图像区域的图像区域特征值依次两两进行比较。
  5. 根据权利要求1所述的图像特征提取设备，其中，所述比较包括将所述至少两个图像区域中的各图像区域的图像区域特征值与基准特征值进行比较。
  6. 根据权利要求5所述的图像特征提取设备，其中，所述基准特征值为所述至少两个图像区域中所选定的基准图像区域的特征值。
  7. 根据权利要求1所述的图像特征提取设备，其中，图像区域的特征值为矢量形式，而图像区域的特征指示符为相应的二进制矢量形式。
  8. 根据权利要求1所述的图像特征提取设备，其中，所述处理电路进一步配置为：
    对图像进行关键点检测，并且
    基于所检测到的各关键点的位置与关键点尺度信息来将图像划分为所述至少两 个图像区域。
  9. 根据权利要求8所述的图像特征提取设备，其中，图像区域是图像中的在所检测到的关键点周围的且具有基于所检测到的关键点尺度信息的尺寸的局部区域。
  10. 根据权利要求8所述的图像特征提取设备，其中，图像区域是图像中的在所检测到的关键点周围的且具有基于所检测到的关键点尺度信息的尺寸的局部区域所包含的子区域。
  11. 根据权利要求8所述的图形特征提取设备，其中，基于高斯-赫米特矩来进行关键点检测，并且其中所述图像区域的特征是基于该图像区域中的像素的高斯-赫米特矩特征而确定的。
  12. 一种图像特征提取方法，包括：
    将图像划分为至少两个图像区域；
    将所述至少两个图像区域中的各图像区域的图像区域特征值与特定特征值进行比较，以基于比较结果来设定图像区域的特征指示符；以及
    组合各图像区域的特征指示符以确定图像的特征。
  13. 根据权利要求12所述的图像特征提取方法，其中，该图像的特征为由各图像区域的特征指示符组成的矢量形式的特征。
  14. 根据权利要求12所述的图像特征提取方法，其中，通过将各图像区域的特征指示符进行级联以得到图像的特征。
  15. 根据权利要求12所述的图像特征提取方法，其中，所述比较包括：将所述至少两个图像区域的图像区域特征值依次两两进行比较。
  16. 根据权利要求12所述的图像特征提取方法，其中，所述比较包括将所述至少两个图像区域中的各图像区域的图像区域特征值与基准特征值进行比较。
  17. 根据权利要求16所述的图像特征提取方法，其中，所述基准特征值为所述至少两个图像区域中所选定的基准图像区域的特征值。
  18. 根据权利要求12所述的图像特征提取方法，其中，图像区域的特征值为矢量形式，而图像区域的特征指示符为相应的二进制矢量形式。
  19. 根据权利要求12所述的图像特征提取方法，进一步包括：
    对图像进行关键点检测，并且
    基于所检测到的各关键点的位置与关键点尺度信息来将图像划分为所述至少两个图像区域。
  20. 根据权利要求19所述的图像特征提取方法，其中，图像区域是图像中的在所检测到的关键点周围的且具有基于所检测到的关键点尺度信息的尺寸的局部区域。
  21. 根据权利要求19所述的图像特征提取方法，其中，图像区域是图像中的在所检测到的关键点周围的且具有基于所检测到的关键点尺度信息的尺寸的局部区域所包含的子区域。
  22. 根据权利要求19所述的图形特征提取方法，其中，基于高斯-赫米特矩来进行关键点检测，并且其中所述图像区域的特征是基于该图像区域中的像素的高斯-赫米特矩特征而确定的。
  23. 一种图像特征提取设备，包括
    至少一个处理器；和
    至少一个存储设备，所述至少一个存储设备在其上存储指令，该指令在由所述至少一个处理器执行时，使所述至少一个处理器执行根据权利要求12-22中任一项所述的方法。
  24. 一种存储指令的存储介质，该指令在由处理器执行时能使得执行根据权利要 求12-22所述的方法。
  25. 一种包括用于执行权利要求12-22所述的方法的部件的图像特征提取装置。
PCT/CN2021/102237 2020-06-30 2021-06-25 图像特征提取方法和设备 WO2022001843A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202180046722.9A CN115956260A (zh) 2020-06-30 2021-06-25 图像特征提取方法和设备

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010611356.4 2020-06-30
CN202010611356.4A CN113869322A (zh) 2020-06-30 2020-06-30 图像特征提取方法和设备

Publications (1)

Publication Number Publication Date
WO2022001843A1 true WO2022001843A1 (zh) 2022-01-06

Family

ID=78981159

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/102237 WO2022001843A1 (zh) 2020-06-30 2021-06-25 图像特征提取方法和设备

Country Status (2)

Country Link
CN (2) CN113869322A (zh)
WO (1) WO2022001843A1 (zh)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105373795A (zh) * 2015-09-18 2016-03-02 中国科学院计算技术研究所 二进制图像特征提取方法及系统
US20160180187A1 (en) * 2014-12-23 2016-06-23 Thomson Licensing Method of generating descriptor for interest point in image and apparatus implementing the same
CN106384127A (zh) * 2016-09-08 2017-02-08 中国科学院计算技术研究所 为图像特征点确定比较点对及二进制描述子的方法及系统
CN110852235A (zh) * 2019-11-05 2020-02-28 长安大学 一种图像特征提取方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160180187A1 (en) * 2014-12-23 2016-06-23 Thomson Licensing Method of generating descriptor for interest point in image and apparatus implementing the same
CN105373795A (zh) * 2015-09-18 2016-03-02 中国科学院计算技术研究所 二进制图像特征提取方法及系统
CN106384127A (zh) * 2016-09-08 2017-02-08 中国科学院计算技术研究所 为图像特征点确定比较点对及二进制描述子的方法及系统
CN110852235A (zh) * 2019-11-05 2020-02-28 长安大学 一种图像特征提取方法

Also Published As

Publication number Publication date
CN113869322A (zh) 2021-12-31
CN115956260A (zh) 2023-04-11

Similar Documents

Publication Publication Date Title
CN110569756B (zh) 人脸识别模型构建方法、识别方法、设备和存储介质
EP3333768A1 (en) Method and apparatus for detecting target
Faraji et al. Face recognition under varying illuminations using logarithmic fractal dimension-based complete eight local directional patterns
US9489566B2 (en) Image recognition apparatus and image recognition method for identifying object
US10565713B2 (en) Image processing apparatus and method
CN106408037B (zh) 图像识别方法及装置
JP6351243B2 (ja) 画像処理装置、画像処理方法
JP6071002B2 (ja) 信頼度取得装置、信頼度取得方法および信頼度取得プログラム
Yin et al. A face anti-spoofing method based on optical flow field
JP2015197708A (ja) オブジェクト識別装置、オブジェクト識別方法及びプログラム
Ismail et al. Efficient enhancement and matching for iris recognition using SURF
CN111222452A (zh) 一种人脸匹配方法、装置、电子设备及可读存储介质
CN111199197B (zh) 一种人脸识别的图像提取方法及处理设备
Wu et al. Privacy leakage of sift features via deep generative model based image reconstruction
WO2022001843A1 (zh) 图像特征提取方法和设备
CN109165551B (zh) 一种自适应加权融合显著性结构张量和lbp特征的表情识别方法
Tsai et al. Recognition of Vehicle License Plates from a Video Sequence.
Walhazi et al. Preprocessing latent-fingerprint images for improving segmentation using morphological snakes
Bae et al. Fingerprint image denoising and inpainting using convolutional neural network
Pang et al. Robust eye center localization through face alignment and invariant isocentric patterns
TWI632509B (zh) 人臉辨識裝置及方法、提升影像辨識率的方法、及電腦可讀儲存介質
Grinchuk et al. Training a multimodal neural network to determine the authenticity of images
Chandra Sekhar et al. Effective splicing localization based on noise level inconsistencies
CN117133022B (zh) 彩色图像掌纹识别方法及装置、设备、存储介质
Kumar et al. Improved navigation for visually challenged with high authentication using a modified sift algorithm

Legal Events

Date Code Title Description
NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21833942

Country of ref document: EP

Kind code of ref document: A1