Object detection device and method based on artificial intelligence

Abstract

The present invention relates to an artificial intelligence-based object detection device and method thereof, comprising: an image acquisition module, configured to acquire a digital image of a target detection object and transmit the digital image to an image preprocessing module; the image preprocessing module, configured to receive the digital image acquired by the image acquisition module, identify and extract the edge of the target detection object in the digital image, perform Fourier transform on the image, perform grayscale adjustment on corresponding pixel points in a low-frequency component area to obtain an enhanced image, and transmit the enhanced image to a regional information statistics module; the regional information statistics module, configured to receive the enhanced image, set a mask matrix field of view, intercept an ROI area, and thereby obtain an array of information of each interval within the ROI area; a data processing optimizer, configured to improve the data processing speed of the enhanced image; and a comprehensive recognition module, configured to integrate feature information of the ROI area, and thereby extract and detect the target detection object.

Description

Technical Field

[0001] The present invention relates to the field of image recognition, and in particular to an object detection device and method based on artificial intelligence. Background Art

[0002] Existing object detection and recognition algorithms typically use the following methods: template matching, back projection, corner detection, and deep learning. Template matching is suitable for rigid objects. Objects with living biological tissues such as humans and animals are subject to significant changes in features due to external conditions, resulting in very high and inaccurate error matching and missed detection rates. Back projection is suitable for searching and matching after knowing the histogram model of an object. However, objects with living biological tissues such as humans and animals are subject to significant changes in features due to external conditions, and the models are not unique. In this case, the accuracy is not good enough and may even lead to false matches. At the same time, the large number of models increases the time consumption. Corner detection works well for objects with significant color and grayscale gradient changes and obvious features, but it is also helpless for objects that lack corner features. Deep learning technology is now widely used for object detection and recognition, but it has relatively high requirements for both software and hardware, which increases costs.

[0003] Chinese patent CN105279484B provides an object detection method and an object detection device. The object detection method includes: mapping at least one frame image in an image sequence to a three-dimensional physical space to obtain the three-dimensional coordinates of each pixel in the image; extracting the foreground area in the image; performing a block operation on the foreground area to obtain a block set; and for each block in the block set, based on the three-dimensional coordinates of a predetermined reference point in the block, detecting the object in the block through a neural network to obtain an object detection result. However, the problem of numerous detection object models, internal image noise interfering with target object detection, and consuming image processing time has not yet been solved. Summary of the Invention

[0004] To this end, the present invention provides an artificial intelligence-based object detection device and method thereof, which can solve the problem that the target detection time is too long due to the large amount of feature information in the image.

[0005] To achieve the above objectives, the present invention provides an artificial intelligence-based object detection device and method thereof, wherein the artificial intelligence-based object detection device comprises:

[0006] An image acquisition module is used to obtain a digital image of the target detection object in the current scene and transmit the digital image to the image preprocessing module;

[0007] The image preprocessing module is used to receive the digital image acquired by the image acquisition module, identify and extract the edge of the target detection object in the digital image, the image preprocessing module sets the set of pixel points located within the u×u neighborhood of each pixel point on the edge of the target detection object and located outside the edge of the target detection object as set A1, and sets the set of pixel points located within the edge of the target detection object as set A2. The image preprocessing module divides the union of the image area formed by set A1 and the image area formed by set A2 into a high-frequency component area and a low-frequency component area according to the discrete degree of the signal frequency in the image, and grayscale adjusts the pixel points in the low-frequency component area that do not meet the low-frequency requirements to obtain an enhanced image. The enhanced image is the union of the image area formed by set A1 after grayscale adjustment and the image area formed by set A2 after grayscale adjustment. The image preprocessing module transmits the enhanced image to the regional information statistics module;

[0008] The regional information statistics module is used to receive the enhanced image transmitted by the image preprocessing module and obtain the image size of the enhanced image and respectively obtain the scanning fields of view of the first mask matrix along the first mask direction and the second mask matrix along the second mask direction for scanning the enhanced image, the regional information statistics module uses the first mask matrix to scan the internal pixels of the enhanced image along the first mask direction to obtain a first interval information array and intercept the ROI region, thereby obtaining image information of the ROI region in the first interval information array, and the regional information statistics module uses the second mask matrix to scan the internal pixels of the enhanced image along the second mask direction to obtain a second interval information array, thereby obtaining image information of the ROI region in the second interval information array, wherein the regional information statistics module sets the first mask direction along the image width direction and sets the second mask direction along the image height direction;

[0009] a data processing optimizer, configured to optimize a data processing flow in a process of scanning pixels inside an enhanced image using the first mask matrix along a first mask direction and the second mask matrix along a second mask direction, so as to increase a data processing speed of the enhanced image;

[0010] The comprehensive identification module is used to integrate the feature information of the ROI area in the first interval array and the second interval array, and then extract and accurately identify the target detection object from various interference objects.

[0011] Furthermore, the image preprocessing module performs grayscale processing on the digital image of the target detection object and obtains the grayscale value Gi of each pixel in the image, i=1, 2, 3, ..., n, wherein G1 is the grayscale value of the first pixel, G2 is the grayscale value of the second pixel, ..., Gn is the grayscale value of the nth pixel, and n is the number of pixels in the digital image of the target detection object. The image preprocessing module establishes an image sub-matrix Mi for the grayscale value of each pixel in the digital image of the target detection object and the grayscale value of each pixel in the 3×3 neighborhood of each pixel, wherein M1 is a first image sub-matrix with the first pixel as the matrix core, M2 is a second image sub-matrix with the second pixel as the matrix core, ..., Mi is an i-th image sub-matrix with the i-th pixel as the matrix core, i= 1,2,…,e, e is the total number of pixels in the enhanced image, wherein the image preprocessing module sets the gray value of the vacant 3×3 neighborhood of the outermost pixel of the enhanced image to 0, and sets the kernel matrix K according to the grayscale amplification degree of each pixel, multiplies each matrix Mi by the kernel matrix K to obtain an enhanced sub-matrix Ji, and adds all element values ​​in each enhanced sub-matrix Ji to obtain a value as the image gradient βi of each pixel, and the image preprocessing module arranges the image gradient βi of each pixel in ascending order and merges the same image gradient values ​​to form a new one-dimensional array R[α1, α2,…, αb], and α1≤α2≤…≤αq, and divides the one-dimensional array R into N intervals, the first interval being (α1, α2,…, α b ), the second interval is (α b+1 , α b+2 ,…,αg),…The Nth interval is (α q-z-1 , α q-z , ..., αq), where b < g < qz-1, b is the number of image gradients in the first interval, g is the sum of the number of image gradients in the first interval and the second interval, z is the sum of the number of image gradients in the first interval to the N-1th interval, q is the number of all image gradients in the enhanced image, and the minimum difference in image gradients in any adjacent interval is greater than or equal to C, C is set as the threshold value of the rate of change of image gradients of adjacent pixel points, and the image preprocessing module determines whether each pixel point is an edge pixel point of the target detection object according to the interval where the image gradient βi of each pixel point is located, where

[0012] When the image gradient βi∈(α q-z-1 , α q-z , ..., αq), the image preprocessing module determines that the i-th pixel is an edge pixel of the target detection object;

[0013] When the image gradient of any pixel in the target detection object image When the pixel points in the Nth interval where the image gradient is located cannot be combined into a continuous smooth curve, the image preprocessing module sequentially selects pixel points with decreasing image gradients as edge pixel points in the order of the N-1th interval, the N-2th interval, ..., the first interval, until they are combined with the pixel points in the Nth interval to form a continuous smooth curve.

[0014] Furthermore, the image preprocessing module sets the closed area formed by the continuous smooth curve composed of each of the edge pixel points as the target detection object, and sets the area composed of pixel points within the m×m neighborhood of each of the edge pixel points and outside the closed area as the vicinity of the target detection object, performs Fourier transform on the grayscale digital image of the target detection object, and obtains the frequency domain of the digital image of the target detection object. The image preprocessing module sets the minimum threshold of the low-frequency component of the digital image of the target detection object to flmin. The image preprocessing module adjusts the grayscale value of the corresponding pixel point in the frequency domain of the digital image of the target detection object according to the comparison result of the regional frequency value fl in the frequency domain of the digital image of the target detection object and the minimum threshold of the low-frequency component flmin, wherein,

[0015] When fl>flmin, the image preprocessing module determines that there is no weak image signal in the digital image of the target detection object and does not adjust each pixel;

[0016] When fl≤flmin, the image preprocessing module determines that there is a weak signal in the digital image of the target detection object, and the image preprocessing module adjusts the grayscale value G of the pixel corresponding to the weak signal area of ​​the digital image of the target detection object to G', so that G'=G×(G / Gavg) G / Gavg ;

[0017] Here, Gavg represents the average grayscale value of each pixel in a 9×9 neighborhood of the pixel.

[0018] Furthermore, when the regional information statistics module receives the enhanced image obtained by the image preprocessing module, the regional information statistics module obtains the size w×h of the enhanced image, obtains the area S1 inside the target detection object, and obtains the area S2 near the target detection object. The regional information statistics module sets the scanning field of view of the first mask matrix when scanning the enhanced image along the first mask direction according to the ratio of the area S1 inside the target detection object to the area S2 near the target detection object, wherein:

[0019] When S1 / S2≤1, the region information statistics module sets the scanning field of view when the first mask matrix scans the enhanced image along the first mask direction to W11×H11=w×(((1-S1 / S2)×1 / 3)×h);

[0020] When S1 / S2>1, the region information statistics module sets the scanning field of view when the first mask matrix scans the enhanced image along the first mask direction to W12×H12=w×(h×1 / 3).

[0021] Furthermore, the regional information statistics module obtains an average signal-to-noise ratio δ of each scanning block obtained by scanning the first mask matrix along the first mask direction, sets δa as the minimum threshold of the image signal-to-noise ratio, and sets a scanning field of view when the second mask matrix scans the enhanced image along the second mask direction based on a comparison result of the average signal-to-noise ratio δ of each scanning block and the minimum threshold of the image signal-to-noise ratio δa, wherein:

[0022] When δ≥δa, the regional information statistics module sets the scanning field of view when the second mask matrix scans the enhanced image along the second mask direction to W21×H21=(1 / 3×w)×h;

[0023] When δ<δa, the region information statistics module sets the scanning field of view when the second mask matrix scans the enhanced image along the second mask direction to W22×H22=(1 / 3×(1+δ / δa)×w)×h.

[0024] Furthermore, the data processing optimizer establishes a plane rectangular coordinate system in the enhanced image plane with the pixel point closest to the grayscale mean of each pixel point in the enhanced image as the coordinate origin. When the mask matrix scans the enhanced image, the data processing optimizer simultaneously performs filtering processing on the enhanced image. The data processing optimizer adopts a Gaussian filtering algorithm with adjustable cutoff frequency. The expression of the Gaussian kernel of the data processing optimizer is: Where (x, y) is the coordinate of the pixel in the image, σ is the standard deviation of the grayscale value of the pixel in the 3×3 neighborhood of the pixel, and the data processing optimizer obtains the adjustment coefficient v according to the field of view size W2×H2 of the mask matrix in the second mask direction, where,

[0025] When W2×H2=W21×H21, the data processing optimizer obtains the adjustment coefficient v=1;

[0026] When W2×H2=W22×H22, the data processing optimizer obtains the adjustment coefficient v=δ / δa.

[0027] Furthermore, the regional information statistics module sets the grayscale intervals to [0, 15], [16, 31], [32, 47], [48, 63], [64, 79], [80, 95], [96, 111], [112, 127], [128, 143], [144, 159], [160, 175], [176, 191], [192, 207], [208, 223], [224, 239], and [240, 255], and the regional information module sets the image grayscale threshold Gp0 according to the number Y of intervals in which each pixel point in the enhanced image is located, wherein,

[0028] When Y>4, the regional information statistics module sets the image grayscale threshold

[0029] When Y≤4, the regional information statistics module sets the image grayscale threshold Gp0=0.5×(Gpmax+Gpmin);

[0030] in, represents the average value of the first grayscale interval where the pixel point in the enhanced image is located, represents the average value of the second grayscale interval where the pixel point in the enhanced image is located, represents the average value of the i-th grayscale interval in which the pixel points in the enhanced image are located, c1 represents the number of pixels in the first grayscale interval, c2 represents the number of pixels in the second grayscale interval, ci represents the number of pixels in the i-th grayscale interval, Gpmax represents the maximum grayscale value of the pixel points in the enhanced image, and Gpmin represents the minimum grayscale value of the pixel points in the enhanced image.

[0031] Furthermore, the region information statistics module performs binarization processing on the enhanced image after grayscale processing to obtain a connected domain. The region information module sets the image grayscale threshold as Gp0, and binarizes the grayscale of each pixel according to the comparison result of the grayscale value Gp of each pixel of the enhanced image after grayscale processing and the image grayscale threshold Gp0, wherein,

[0032] When Gp≤Gp0, the regional information statistics module adjusts the grayscale of the corresponding pixel to 0;

[0033] When Gp>Gp0, the region information statistics module adjusts the grayscale of the corresponding pixel to 255.

[0034] Furthermore, the regional information statistics module selects the connected domain with the largest area Ymax, and obtains the area of ​​each connected domain and the pixel points located on the connected domain outline. The regional information statistics module obtains the minimum distance lk between the connected domain Ymax outline point and the connected domain outline, k=1, 2, ...n, n is the number of connected domains, and determines whether each connected domain is valid based on the area Sk of each connected domain and the minimum distance lk between the connected domain Ymax outline point and the connected domain outline, wherein,

[0035] When Sk≤1 / 50×Smax, the region information statistics module determines that the connected domain is invalid and does not belong to the ROI region;

[0036] When 1 / 50×Smax<Sk≤1 / 30×Smax and lk>dmax, the region information statistics module determines that the connected domain is invalid and does not belong to the ROI region;

[0037] When 1 / 50×Smax<Sk≤1 / 30×Smax and lk≤dmax, the region information statistics module determines that the connected domain is valid and belongs to the ROI region;

[0038] When Sk>1 / 30×Smax, the region information statistics module determines that the connected domain is valid and belongs to the ROI region;

[0039] Among them, Smax represents the area of ​​the connected domain Ymax, and dmax represents the straight-line distance between the two pixels with the farthest straight-line distance on the contour of the connected domain Ymax.

[0040] Furthermore, the regional information statistics module extracts each pixel point on the effective connected domain contour line, and selects the upper left vertex, the upper right vertex, the lower left vertex and the lower right vertex from all the pixel points on the effective connected domain contour line. The regional information statistics module connects the upper left vertex, the upper right vertex, the lower left vertex and the lower right vertex in a straight line so that each vertex is connected to form a quadrilateral. The regional information statistics module intercepts the ROI area according to the coverage area of ​​the quadrilateral, wherein,

[0041] When the quadrilateral covers all valid connected domains, the regional information statistics module intercepts the quadrilateral as the ROI area;

[0042] When the quadrilateral cannot cover all valid connected domains, the region information statistics module intercepts a bounding rectangle that can cover all valid connected domains as the ROI region.

[0043] Compared with the existing technology, the beneficial effect of the present invention is that it can save the generation and extraction of templates during template matching, and does not require skilled software and hardware as demanding as deep learning, and at the same time has relatively high reliability; the present invention has relatively profound significance for the identification and detection of biological tissue objects, freeing technical R&D personnel from complex feature extraction methods, allowing more time and energy to be invested in the research of core issues. In addition, in terms of time efficiency, it can save a lot of detection time and improve detection efficiency.

[0044] In particular, the present invention sets an image preprocessing module to preprocess the image, enhances the relatively weak signals near and inside the target image to obtain an enhanced image, which is convenient for subsequent detection and identification of objects; the present invention sets a regional information statistics module, and uses a mask matrix to scan the target image, so that the values ​​in the neighborhood of each pixel point can be weighted averaged during the scanning process, thereby reducing the omission of detailed feature information in the target image, and making the image information in the obtained interval information array referenceable, which is conducive to the detection and identification of target objects; the present invention sets a data processing optimizer to optimize the data processing flow in the mask matrix, speed up the scanning speed of the mask matrix, and improve the efficiency of object detection; the present invention sets a comprehensive identification module, which can comprehensively analyze the interval information arrays obtained by two mask matrix scans, and can reduce the errors in object detection and identification results.

[0045] In particular, the present invention amplifies the data in the grayscale matrix formed by each pixel point in the image during the preprocessing process based on the kernel matrix algorithm, which is conducive to the accurate edge extraction of the target detection object by the artificial intelligence-based object detection device and method thereof; the present invention determines whether the grayscale jump occurs between adjacent pixels by the grayscale gradient change of each pixel point, and preferentially selects pixels with higher image gradient to improve the accuracy of image edge extraction. Selecting pixels as edge contour points in the order of decreasing image gradient can effectively avoid the extracted edge contour line from being unable to fit into a continuous smooth curve, making subsequent recognition difficult.

[0046] In particular, the present invention performs a Fourier transform on a digital image to obtain the image's high-frequency and low-frequency components. The lower-frequency signal regions within the low-frequency components are then used as signal enhancement areas. The high and low frequencies of an image are a measure of the intensity variation between image locations. Pixel grayscale values ​​in the low-frequency signal region do not change significantly, making detailed feature information easily lost during image detection and recognition. Therefore, adjusting the grayscale of pixels in the low-frequency signal region improves image contrast, making detailed feature information easier to detect and recognize.

[0047] In particular, the present invention sets the field of view size of the mask matrix according to the area of ​​the target detection object itself in the enhanced image and the area near the target detection object in the enhanced image. When the area of ​​the target detection object is smaller than its background area, selecting a smaller matrix field of view under a constant matrix condition can improve the spatial resolution and avoid losing important image feature information; when the area of ​​the target detection object is larger than its background area, selecting a larger matrix field of view can speed up the image data processing speed and improve the recognition efficiency of the target detection object.

[0048] In particular, the present invention obtains the signal-to-noise ratio of the enhanced image of the target object, estimates the effective information content and noise content in the enhanced image, and when the noise content is low, selecting a smaller matrix field of view can obtain more image feature information while ensuring that the noise interference in the image is small; when the noise content is high, in order to avoid excessive noise interference, selecting a larger matrix field of view can reduce the interference of noise on the image recognition result while screening the more important feature information of the image, thereby reducing the error of the target detection object recognition result.

[0049] In particular, the present invention uses a data processing optimizer based on Gaussian filtering to process image noise, which can effectively suppress noise that obeys the normal distribution. The present invention adjusts the cutoff frequency of the data processing optimizer by setting an adjustment coefficient. When the mask matrix field of view is small, selecting a larger adjustment coefficient can improve the denoising degree of the image, reduce the impact of noise on the target detection object, and increase the scanning speed of the mask matrix; when the mask matrix field of view is large, selecting a smaller adjustment coefficient can avoid mistakenly treating image detail feature information as noise, thereby avoiding the loss of important feature information of the image.

[0050] In particular, the present invention estimates the contrast of the enhanced image based on the number of grayscale intervals where each pixel point is located. When the number of grayscale intervals where each pixel point is located is large, it can be determined that the contrast of the enhanced image is high and the grayscale values ​​are large. The grayscale weighted average method can make the grayscale threshold positioning more accurate, which is beneficial to subsequent image processing; when the number of grayscale intervals where each pixel point is located is small, it can be determined that the image contrast is small. Selecting the image grayscale median as the grayscale threshold can effectively divide the image.

[0051] In particular, the present invention performs polarization processing on the grayscale value of each pixel, so that only black and white blocks exist in the grayscale enhanced image. The collective properties of the image are only related to the positions of points with pixel values ​​of 0 or 255, and no longer involve multi-level values ​​of pixels, which makes processing simple. In addition, the amount of data processing and compression is small, and the outline of the target of interest can be highlighted, which is conducive to the regional information statistics module to identify closed figures and then accurately extract connected domains.

[0052] In particular, the present invention divides the effective connected domains and invalid connected domains according to the area size relationship and position relationship of each connected domain. Since part of the noise in the image will also be enhanced when the original image signal is enhanced, there will be interference feature information in the enhanced image. By comparing the areas of each connected domain and analyzing the position relationship, part of the interference feature information can be effectively eliminated, reducing the adverse effects of the interference feature information on the detection of the target object. At the same time, reducing the amount of interference feature information helps to reduce the time for target object detection and identification, and improve the detection efficiency of the artificial intelligence-based object detection device and method thereof.

[0053] In particular, the present invention takes the effective connected domain as a subset of the ROI area, and reduces the complement range of the effective connected domain in the ROI area as much as possible, that is, minimizes the background area of ​​the effective connected domain. The present invention preferentially selects the four pixel points with the farthest distance from the entire effective connected domain module, that is, the upper left vertex, the lower left vertex, the upper right vertex and the lower right vertex, and connects these four pixel points into a quadrilateral. When the quadrilateral area can cover the effective connected domain, using the quadrilateral area as the ROI area can reduce the background area in the ROI area, thereby reducing the amount of image information processing; when the quadrilateral area cannot cover the effective connected domain, another circumscribed rectangle is intercepted as the ROI area, minimizing the invalid image information that needs to be processed while ensuring that no image information is lost. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a schematic diagram showing the connections between the various systems within an artificial intelligence-based object detection device according to an embodiment of the invention;

[0055] Figure 2 A first mask direction of the mask matrix of an embodiment of the invention;

[0056] Figure 3 This is the second mask direction of the mask matrix according to the embodiment of the invention. DETAILED DESCRIPTION

[0057] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are merely used to explain the present invention and are not intended to limit the present invention.

[0058] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0059] It should be noted that, in the description of the present invention, terms such as "up", "down", "left", "right", "inside", and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on the present invention.

[0060] Furthermore, it should be noted that, in the description of the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0061] See also Figure 1 FIG. 1 is a schematic diagram showing the connection of various systems within an artificial intelligence-based object detection device according to an embodiment of the present invention. The artificial intelligence-based object detection device includes:

[0062] An image acquisition module is used to obtain a digital image of the target detection object in the current scene and transmit the digital image to the image preprocessing module;

[0063] The image preprocessing module is used to receive the digital image acquired by the image acquisition module, identify and extract the edge of the target detection object in the digital image, the image preprocessing module sets the set of pixel points located within the u×u neighborhood of each pixel point on the edge of the target detection object and located outside the edge of the target detection object as set A1, and sets the set of pixel points located within the edge of the target detection object as set A2. The image preprocessing module divides the union of the image area formed by set A1 and the image area formed by set A2 into a high-frequency component area and a low-frequency component area according to the discrete degree of the signal frequency in the image, and grayscale adjusts the pixel points in the low-frequency component area that do not meet the low-frequency requirements to obtain an enhanced image. The enhanced image is the union of the image area formed by set A1 after grayscale adjustment and the image area formed by set A2 after grayscale adjustment. The image preprocessing module transmits the enhanced image to the regional information statistics module;

[0064] The regional information statistics module is used to receive the enhanced image transmitted by the image preprocessing module and obtain the image size of the enhanced image and respectively obtain the scanning fields of view of the first mask matrix along the first mask direction and the second mask matrix along the second mask direction for scanning the enhanced image, the regional information statistics module uses the first mask matrix to scan the internal pixels of the enhanced image along the first mask direction to obtain a first interval information array and intercept the ROI region, thereby obtaining image information of the ROI region in the first interval information array, and the regional information statistics module uses the second mask matrix to scan the internal pixels of the enhanced image along the second mask direction to obtain a second interval information array, thereby obtaining image information of the ROI region in the second interval information array, wherein the regional information statistics module sets the first mask direction along the image width direction and sets the second mask direction along the image height direction;

[0065] a data processing optimizer, configured to optimize a data processing flow in a process of scanning pixels inside an enhanced image using the first mask matrix along a first mask direction and the second mask matrix along a second mask direction, so as to increase a data processing speed of the enhanced image;

[0066] The comprehensive identification module is used to integrate the feature information of the ROI area in the first interval array and the second interval array, and then extract and accurately identify the target detection object from various interference objects.

[0067] See also Figure 2 and Figure 3 As shown, they are respectively the first mask direction and the second mask direction of the mask matrix of the embodiment of the invention.

[0068] Specifically, the present invention sets an image preprocessing module to preprocess the image, enhances the relatively weak signals near and inside the target image to obtain an enhanced image, which is convenient for subsequent detection and recognition of objects; the present invention sets a regional information statistics module, and uses a mask matrix to scan the target image, so that the values ​​in the neighborhood of each pixel point can be weighted averaged during the scanning process, thereby reducing the omission of detailed feature information in the target image, and making the image information in the obtained interval information array referenceable, which is conducive to the detection and recognition of target objects; the present invention sets a data processing optimizer to optimize the data processing flow in the mask matrix, speed up the scanning speed of the mask matrix, and improve the efficiency of object detection; the present invention sets a comprehensive recognition module, which can comprehensively analyze the interval information array obtained by two mask matrix scans, and can reduce the error of object detection and recognition results.

[0069] The image preprocessing module performs grayscale processing on the digital image of the target detection object and obtains the grayscale value Gi of each pixel in the image, i=1, 2, 3, ..., n, wherein G1 is the grayscale value of the first pixel, G2 is the grayscale value of the second pixel, ..., Gn is the grayscale value of the nth pixel, and n is the number of pixels in the digital image of the target detection object. The image preprocessing module establishes an image sub-matrix Mi for the grayscale value of each pixel in the digital image of the target detection object and the grayscale value of each pixel in the 3×3 neighborhood of each pixel, wherein M1 is a first image sub-matrix with the first pixel as the matrix core, M2 is a second image sub-matrix with the second pixel as the matrix core, ..., Mi is an i-th image sub-matrix with the i-th pixel as the matrix core, i=1, 2 ,…,e, e is the total number of pixels in the enhanced image, wherein the image preprocessing module sets the gray value of the vacant 3×3 neighborhood of the outermost pixel of the enhanced image to 0, and sets the kernel matrix K according to the grayscale amplification degree of each pixel, multiplies each matrix Mi by the kernel matrix K to obtain an enhanced sub-matrix Ji, and adds all element values ​​in each enhanced sub-matrix Ji to obtain a value as the image gradient βi of each pixel, and the image preprocessing module arranges the image gradient βi of each pixel in ascending order and merges the same image gradient values ​​to form a new one-dimensional array R[α1, α2,…, αb], and α1≤α2≤…≤αq, and divides the one-dimensional array R into N intervals, the first interval is (α1, α2,…, α b ), the second interval is (α b+1 , α b+2 ,…,αg),…The Nth interval is (α q-z-1 , α q-z , ..., αq), where b < g < qz-1, b is the number of image gradients in the first interval, g is the sum of the number of image gradients in the first interval and the second interval, z is the sum of the number of image gradients in the first interval to the N-1th interval, q is the number of all image gradients in the enhanced image, and the minimum difference in image gradients in any adjacent interval is greater than or equal to C, C is set as the threshold value of the rate of change of image gradients of adjacent pixel points, and the image preprocessing module determines whether each pixel point is an edge pixel point of the target detection object according to the interval where the image gradient βi of each pixel point is located, where

[0070] When the image gradient βi∈(α q-z-1 , α q-z , ..., αq), the image preprocessing module determines that the i-th pixel is an edge pixel of the target detection object;

[0071] When the image gradient of any pixel in the target detection object image When the pixel points in the Nth interval where the image gradient is located cannot be combined into a continuous smooth curve, the image preprocessing module sequentially selects pixel points with decreasing image gradients as edge pixel points in the order of the N-1th interval, the N-2th interval, ..., the first interval, until they are combined with the pixel points in the Nth interval to form a continuous smooth curve.

[0072] Specifically, the present invention does not limit the core matrix. The present invention does not limit the threshold value of the gradient change rate of adjacent pixel points. The threshold value of the gradient change rate of adjacent pixel points is a critical value of the image gradient jump; the minimum difference of the image gradient in any adjacent interval of the present invention is the difference between the maximum value of the previous interval and the minimum value of the next interval.

[0073] Specifically, the present invention amplifies the data in the grayscale matrix formed by each pixel point in the image during the preprocessing process based on the kernel matrix algorithm, which is conducive to the artificial intelligence-based object detection device and method to accurately extract the edge of the target detection object; the present invention determines whether the grayscale jump occurs between adjacent pixels by the grayscale gradient change of each pixel point, and preferentially selects pixels with higher image gradient to improve the accuracy of image edge extraction. Selecting pixels as edge contour points in the order of decreasing image gradient can effectively avoid the extracted edge contour line from being unable to fit into a continuous smooth curve, making subsequent recognition difficult.

[0074] The image preprocessing module sets the closed area formed by the continuous smooth curve composed of each of the edge pixel points as the target detection object, and sets the area composed of pixel points within the m×m neighborhood of each of the edge pixel points and outside the closed area as the vicinity of the target detection object, performs Fourier transform on the grayscale digital image of the target detection object, and obtains the frequency domain of the digital image of the target detection object. The image preprocessing module sets the minimum threshold of the low-frequency component of the digital image of the target detection object to flmin. The image preprocessing module adjusts the grayscale value of the corresponding pixel point in the frequency domain of the digital image of the target detection object based on the comparison result of the regional frequency value fl in the frequency domain of the digital image of the target detection object and the minimum threshold of the low-frequency component flmin, wherein,

[0075] When fl>flmin, the image preprocessing module determines that there is no weak image signal in the digital image of the target detection object and does not adjust each pixel;

[0076] When fl≤flmin, the image preprocessing module determines that there is a weak signal in the digital image of the target detection object, and the image preprocessing module adjusts the grayscale value G of the pixel corresponding to the weak signal area of ​​the digital image of the target detection object to G', so that G'=G×(G / Gavg) G / Gavg ;

[0077] Here, Gavg represents the average grayscale value of each pixel in a 9×9 neighborhood of the pixel.

[0078] Specifically, the present invention performs a Fourier transform on a digital image to obtain the high-frequency and low-frequency components of the image. The low-frequency signal regions within the low-frequency components are then used as signal enhancement regions. The high and low frequencies of an image are a measure of the intensity variation between different image locations. The grayscale values ​​of pixels in the low-frequency signal region do not change significantly, making detailed feature information easily lost during image detection and recognition. Therefore, adjusting the grayscale values ​​of pixels in the low-frequency signal region improves image contrast, making detailed feature information easier to detect and recognize.

[0079] When the regional information statistics module receives the enhanced image obtained by the image preprocessing module, the regional information statistics module obtains the size w×h of the enhanced image, obtains the area S1 within the target detection object, and obtains the area S2 near the target detection object. The regional information statistics module sets the scanning field of view of the first mask matrix when scanning the enhanced image along the first mask direction according to the ratio of the area S1 within the target detection object to the area S2 near the target detection object, wherein

[0080] When S1 / S2≤1, the region information statistics module sets the scanning field of view when the first mask matrix scans the enhanced image along the first mask direction to W11×H11=w×(((1-S1 / S2)×1 / 3)×h);

[0081] When S1 / S2>1, the region information statistics module sets the scanning field of view when the first mask matrix scans the enhanced image along the first mask direction to W12×H12=w×(h×1 / 3).

[0082] Specifically, the present invention sets the field of view size of the mask matrix according to the area of ​​the target detection object itself in the enhanced image and the area near the target detection object in the enhanced image. When the area of ​​the target detection object is smaller than its background area, selecting a smaller matrix field of view under a constant matrix condition can improve the spatial resolution and avoid losing important image feature information; when the area of ​​the target detection object is larger than its background area, selecting a larger matrix field of view can speed up the image data processing speed and improve the recognition efficiency of the target detection object.

[0083] The regional information statistics module obtains the average signal-to-noise ratio δ of each scanning block obtained by scanning the first mask matrix along the first mask direction, sets δa as the minimum threshold of the image signal-to-noise ratio, and sets the scanning field of view when the second mask matrix scans the enhanced image along the second mask direction according to the comparison result of the average signal-to-noise ratio δ of each scanning block and the minimum threshold of the image signal-to-noise ratio δa, wherein:

[0084] When δ≥δa, the regional information statistics module sets the scanning field of view when the second mask matrix scans the enhanced image along the second mask direction to W21×H21=(1 / 3×w)×h;

[0085] When δ<δa, the region information statistics module sets the scanning field of view when the second mask matrix scans the enhanced image along the second mask direction to W22×H22=(1 / 3×(1+δ / δa)×w)×h.

[0086] Specifically, the present invention obtains the signal-to-noise ratio of the enhanced image of the target object, estimates the effective information content and noise content in the enhanced image, and when the noise content is low, selecting a smaller matrix field of view can obtain more image feature information while ensuring that the noise interference in the image is small; when the noise content is high, in order to avoid excessive noise interference, selecting a larger matrix field of view can reduce the interference of noise on the image recognition results while screening the more important feature information of the image, thereby reducing the error of the target detection object recognition result.

[0087] The data processing optimizer establishes a plane rectangular coordinate system on the enhanced image plane with the pixel point closest to the grayscale mean of each pixel point of the enhanced image as the coordinate origin. When the mask matrix scans the enhanced image, the data processing optimizer simultaneously performs filtering processing on the enhanced image. The data processing optimizer adopts a Gaussian filtering algorithm with an adjustable cutoff frequency. The expression of the Gaussian kernel of the data processing optimizer is h(x, Where (x, y) is the coordinate of the pixel in the image, σ is the standard deviation of the grayscale value of the pixel in the 3×3 neighborhood of the pixel, and the data processing optimizer obtains the adjustment coefficient v according to the field of view size W2×H2 of the mask matrix in the second mask direction, where,

[0088] When W2×H2=W21×H21, the data processing optimizer obtains the adjustment coefficient v=1;

[0089] When W2×H2=W22×H22, the data processing optimizer obtains the adjustment coefficient v=δ / δa.

[0090] Specifically, the present invention uses a data processing optimizer based on Gaussian filtering to process image noise, which can effectively suppress noise that obeys the normal distribution. The present invention adjusts the cutoff frequency of the data processing optimizer by setting an adjustment coefficient. When the mask matrix field of view is small, selecting a larger adjustment coefficient can improve the denoising degree of the image, reduce the impact of noise on the target detection object, and increase the scanning speed of the mask matrix; when the mask matrix field of view is large, selecting a smaller adjustment coefficient can avoid mistakenly treating image detail feature information as noise, thereby avoiding the loss of important feature information of the image.

[0091] The regional information statistics module sets the grayscale intervals to [0, 15], [16, 31], [32, 47], [48, 63], [64, 79], [80, 95], [96, 111], [112, 127], [128, 143], [144, 159], [160, 175], [176, 191], [192, 207], [208, 223], [224, 239], and [240, 255]. The regional information module sets the image grayscale threshold Gp0 according to the number Y of intervals where each pixel point in the enhanced image is located, wherein

[0092] When Y>4, the regional information statistics module sets the image grayscale threshold

[0093] When Y≤4, the regional information statistics module sets the image grayscale threshold Gp0=0.5×(Gpmax+Gpmin);

[0094] in, represents the average value of the first grayscale interval where the pixel point in the enhanced image is located, represents the average value of the second grayscale interval where the pixel point in the enhanced image is located, represents the average value of the i-th grayscale interval in which the pixel points in the enhanced image are located, c1 represents the number of pixels in the first grayscale interval, c2 represents the number of pixels in the second grayscale interval, ci represents the number of pixels in the i-th grayscale interval, Gpmax represents the maximum grayscale value of the pixel points in the enhanced image, and Gpmin represents the minimum grayscale value of the pixel points in the enhanced image.

[0095] Specifically, the present invention estimates the contrast of the enhanced image based on the number of grayscale intervals where each pixel point is located. When the number of grayscale intervals where each pixel point is located is large, it can be determined that the contrast of the enhanced image is high and the grayscale values ​​are large. The grayscale weighted average method can make the grayscale threshold positioning more accurate, which is beneficial to subsequent image processing; when the number of grayscale intervals where each pixel point is located is small, it can be determined that the image contrast is small. Selecting the image grayscale median as the grayscale threshold can effectively divide the image.

[0096] The regional information statistics module performs binarization processing on the enhanced image after grayscale processing to obtain the connected domain. The regional information module sets the image grayscale threshold to Gp0, and binarizes the grayscale of each pixel according to the comparison result of the grayscale value Gp of each pixel in the enhanced image after grayscale processing and the image grayscale threshold Gp0, wherein,

[0097] When Gp≤Gp0, the regional information statistics module adjusts the grayscale of the corresponding pixel to 0;

[0098] When Gp>Gp0, the region information statistics module adjusts the grayscale of the corresponding pixel to 255.

[0099] Specifically, the present invention performs polarization processing on the grayscale value of each pixel point, so that only black and white color blocks exist in the grayscale enhanced image. The collective properties of the image are only related to the positions of points with pixel values ​​0 or 255, and no longer involve multi-level values ​​of pixels, which makes processing simpler. In addition, the amount of data processing and compression is small, and the outline of the target of interest can be highlighted, which is beneficial for the regional information statistics module to identify closed figures and then accurately extract connected domains.

[0100] The regional information statistics module selects the connected domain with the largest area Ymax, and obtains the area of ​​each connected domain and the pixel points located on the connected domain outline. The regional information statistics module obtains the minimum distance lk between the connected domain Ymax outline point and the connected domain outline, k=1,2,…n, n is the number of connected domains, and determines whether each connected domain is valid based on the area Sk of each connected domain and the minimum distance lk between the connected domain Ymax outline point and the connected domain outline, wherein,

[0101] When Sk≤1 / 50×Smax, the region information statistics module determines that the connected domain is invalid and does not belong to the ROI region;

[0102] When 1 / 50×Smax<Sk≤1 / 30×Smax and lk>dmax, the region information statistics module determines that the connected domain is invalid and does not belong to the ROI region;

[0103] When 1 / 50×Smax<Sk≤1 / 30×Smax and lk≤dmax, the region information statistics module determines that the connected domain is valid and belongs to the ROI region;

[0104] When Sk>1 / 30×Smax, the region information statistics module determines that the connected domain is valid and belongs to the ROI region;

[0105] Among them, Smax represents the area of ​​the connected domain Ymax, and dmax represents the straight-line distance between the two pixels with the farthest straight-line distance on the contour of the connected domain Ymax.

[0106] Specifically, the present invention divides the effective connected domains and invalid connected domains according to the area size relationship and position relationship of each connected domain. Since part of the noise in the image will be enhanced when the original image signal is enhanced, there will be interference feature information in the enhanced image. By comparing the areas of each connected domain and analyzing the position relationship, part of the interference feature information can be effectively eliminated, reducing the adverse effects of the interference feature information on the detection of the target object. At the same time, reducing the amount of interference feature information helps to reduce the time for target object detection and identification, thereby improving the detection efficiency of the artificial intelligence-based object detection device and method thereof.

[0107] The regional information statistics module extracts each pixel point on the effective connected domain contour line, and selects the upper left vertex, the upper right vertex, the lower left vertex and the lower right vertex from all the pixel points on the effective connected domain contour line. The regional information statistics module connects the upper left vertex, the upper right vertex, the lower left vertex and the lower right vertex in a straight line so that each vertex is connected to form a quadrilateral. The regional information statistics module intercepts the ROI area according to the coverage area of ​​the quadrilateral, wherein,

[0108] When the quadrilateral covers all valid connected domains, the regional information statistics module intercepts the quadrilateral as the ROI area;

[0109] When the quadrilateral cannot cover all valid connected domains, the region information statistics module intercepts a bounding rectangle that can cover all valid connected domains as the ROI region.

[0110] Specifically, the present invention takes the effective connected domain as a subset of the ROI area, and reduces the complement range of the effective connected domain in the ROI area as much as possible, that is, minimizes the background area of ​​the effective connected domain. The present invention preferentially selects the four pixel points with the farthest distance from the entire effective connected domain module, that is, the upper left vertex, the lower left vertex, the upper right vertex and the lower right vertex, and connects these four pixel points into a quadrilateral. When the quadrilateral area can cover the effective connected domain, using the quadrilateral area as the ROI area can reduce the background area within the ROI area, thereby reducing the amount of image information processing; when the quadrilateral area cannot cover the effective connected domain, another circumscribed rectangle is intercepted as the ROI area to minimize the invalid image information that needs to be processed without losing image information.

[0111] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

[0112] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. An object detection device based on artificial intelligence, characterized in that: include: An image acquisition module is used to obtain a digital image of the target detection object in the current scene and transmit the digital image to the image preprocessing module; The image preprocessing module is used to receive the digital image acquired by the image acquisition module, identify and extract the edge of the target detection object in the digital image, the image preprocessing module sets the set of pixel points located within the u×u neighborhood of each pixel point on the edge of the target detection object and located outside the edge of the target detection object as set A1, and sets the set of pixel points located within the edge of the target detection object as set A2. The image preprocessing module divides the union of the image area formed by set A1 and the image area formed by set A2 into a high-frequency component area and a low-frequency component area according to the discrete degree of the signal frequency in the image, and grayscale adjusts the pixel points in the low-frequency component area that do not meet the low-frequency requirements to obtain an enhanced image. The enhanced image is the union of the image area formed by set A1 after grayscale adjustment and the image area formed by set A2 after grayscale adjustment. The image preprocessing module transmits the enhanced image to the regional information statistics module; The regional information statistics module is used to receive the enhanced image transmitted by the image preprocessing module and obtain the image size of the enhanced image and respectively obtain the scanning fields of view of the first mask matrix along the first mask direction and the second mask matrix along the second mask direction for scanning the enhanced image, the regional information statistics module uses the first mask matrix to scan the internal pixels of the enhanced image along the first mask direction to obtain a first interval information array and intercept the ROI region, thereby obtaining image information of the ROI region in the first interval information array, and the regional information statistics module uses the second mask matrix to scan the internal pixels of the enhanced image along the second mask direction to obtain a second interval information array, thereby obtaining image information of the ROI region in the second interval information array, wherein the regional information statistics module sets the first mask direction along the image width direction and sets the second mask direction along the image height direction; a data processing optimizer, configured to optimize a data processing flow in a process of scanning pixels inside an enhanced image using the first mask matrix along a first mask direction and the second mask matrix along a second mask direction, so as to increase a data processing speed of the enhanced image; The comprehensive identification module is used to integrate the feature information of the ROI area in the first interval array and the second interval array, and then extract and accurately identify the target detection object from various interference objects.

2. The artificial intelligence-based object detection device according to claim 1, characterized in that The image preprocessing module performs grayscale processing on the digital image of the target detection object and obtains the grayscale value Gi of each pixel in the image, i=1, 2, 3, ..., n, wherein G1 is the grayscale value of the first pixel, G2 is the grayscale value of the second pixel, ..., Gn is the grayscale value of the nth pixel, and n is the number of pixels in the digital image of the target detection object. The image preprocessing module establishes an image sub-matrix Mi for the grayscale value of each pixel in the digital image of the target detection object and the grayscale value of each pixel in the 3×3 neighborhood of each pixel, wherein M1 is a first image sub-matrix with the first pixel as the matrix core, M2 is a second image sub-matrix with the second pixel as the matrix core, ..., Mi is a matrix with the i-th pixel as the matrix core. The i-th image sub-matrix of the kernel, i=1,2,…,e, e is the total number of pixels in the enhanced image, wherein the image preprocessing module sets the gray value of the vacant 3×3 neighborhood of the outermost pixel of the enhanced image to 0, and sets the kernel matrix K according to the grayscale amplification degree of each pixel, multiplies each of the image sub-matrices Mi by the kernel matrix K to obtain an enhanced sub-matrix Ji, and sums all the element values ​​in each enhanced sub-matrix Ji to obtain a value as the image gradient βi of each pixel, and the image preprocessing module arranges the image gradient βi of each pixel in ascending order and merges the same image gradient values ​​to form a new one-dimensional array R[α1, α2,…, αb], and α1≤α2≤…≤α q , divide the one-dimensional array R into N intervals, the first interval is (α1, α2, ..., α b ), the second interval is (α b+1 , α b+2 ,…,α g ), ... the Nth interval is (α q-z-1 , α q-z ,……,α q ), wherein b<g<qz-1, b is the number of image gradients in the first interval, g is the sum of the number of image gradients in the first interval and the second interval, z is the sum of the number of image gradients in the first interval to the N-1th interval, q is the number of all image gradients in the enhanced image, and the minimum difference of the image gradients in any adjacent interval is greater than or equal to C, and C is set as the threshold value of the change rate of the image gradients of adjacent pixels. The image preprocessing module determines whether each pixel is an edge pixel of the target detection object according to the interval where the image gradient βi of each pixel is located, wherein, When the image gradient βi∈(α q-z-1 , α q-z ,……,α q ), the image preprocessing module determines that the i-th pixel point is an edge pixel point of the target detection object; When the image gradient of any pixel in the target detection object image When the pixel points in the Nth interval where the image gradient is located cannot be combined into a continuous smooth curve, the image preprocessing module sequentially selects pixel points with decreasing image gradients as edge pixel points in the order of the N-1th interval, the N-2th interval, ..., the first interval, until they are combined with the pixel points in the Nth interval to form a continuous smooth curve.

3. The artificial intelligence-based object detection device according to claim 2, characterized in that: The image preprocessing module sets the closed area formed by the continuous smooth curve composed of each of the edge pixel points as the target detection object, and sets the area composed of pixel points within the m×m neighborhood of each of the edge pixel points and outside the closed area as the vicinity of the target detection object, performs Fourier transform on the grayscale digital image of the target detection object, and obtains the frequency domain of the digital image of the target detection object. The image preprocessing module sets the minimum threshold of the low-frequency component of the digital image of the target detection object to flmin. The image preprocessing module adjusts the grayscale value of the corresponding pixel point in the frequency domain of the digital image of the target detection object based on the comparison result of the regional frequency value fl in the frequency domain of the digital image of the target detection object and the minimum threshold of the low-frequency component flmin, wherein, When fl>flmin, the image preprocessing module determines that there is no weak image signal in the digital image of the target detection object and does not adjust each pixel; When fl≤flmin, the image preprocessing module determines that there is a weak image signal in the digital image of the target detection object, and the image preprocessing module adjusts the grayscale value G of the pixel corresponding to the weak signal area of ​​the digital image of the target detection object to G', so that G'=G×(G / Gavg)G / Gavg; Here, Gavg represents the average grayscale value of each pixel in a 9×9 neighborhood of the pixel.

4. The artificial intelligence-based object detection device according to claim 3, characterized in that: When the regional information statistics module receives the enhanced image obtained by the image preprocessing module, the regional information statistics module obtains the size w×h of the enhanced image, obtains the area S1 within the target detection object, and obtains the area S2 near the target detection object. The regional information statistics module sets the scanning field of view of the first mask matrix when scanning the enhanced image along the first mask direction according to the ratio of the area S1 within the target detection object to the area S2 near the target detection object, wherein When S1 / S2≤1, the region information statistics module sets the scanning field of view when the first mask matrix scans the enhanced image along the first mask direction to W11×H11=w×(((1-S1 / S2)×1 / 3)×h); When S1 / S2>1, the region information statistics module sets the scanning field of view when the first mask matrix scans the enhanced image along the first mask direction to W12×H12=w×(h×1 / 3).

5. The artificial intelligence-based object detection device according to claim 4, characterized in that: The regional information statistics module obtains the average signal-to-noise ratio δ of each scanning block obtained by scanning the first mask matrix along the first mask direction, sets δa as the minimum threshold of the image signal-to-noise ratio, and sets the scanning field of view when the second mask matrix scans the enhanced image along the second mask direction according to the comparison result of the average signal-to-noise ratio δ of each scanning block and the minimum threshold of the image signal-to-noise ratio δa, wherein: When δ≥δa, the regional information statistics module sets the scanning field of view when the second mask matrix scans the enhanced image along the second mask direction to W21×H21=(1 / 3×w)×h; When δ<δa, the region information statistics module sets the scanning field of view when the second mask matrix scans the enhanced image along the second mask direction to W22×H22=(1 / 3×(1+δ / δa)×w)×h.

6. The artificial intelligence-based object detection device according to claim 5, characterized in that: The data processing optimizer establishes a plane rectangular coordinate system in the enhanced image plane with the pixel point closest to the grayscale mean of each pixel point of the enhanced image as the coordinate origin. When the mask matrix scans the enhanced image, the data processing optimizer simultaneously performs filtering processing on the enhanced image. The data processing optimizer adopts a Gaussian filtering algorithm with adjustable cutoff frequency. The expression of the Gaussian kernel of the data processing optimizer is: Where (x, y) is the coordinate of the pixel in the image, σ is the standard deviation of the grayscale value of the pixel in the 3×3 neighborhood of the pixel, and the data processing optimizer obtains the adjustment coefficient v according to the field of view size W2×H2 of the mask matrix in the second mask direction, where, When W2×H2=W21×H21, the data processing optimizer obtains the adjustment coefficient v=1; When W2×H2=W22×H22, the data processing optimizer obtains the adjustment coefficient v=δ / δa.

7. The artificial intelligence-based object detection device according to claim 5, characterized in that: The regional information statistics module sets the grayscale intervals to [0, 15], [16, 31], [32, 47], [48, 63], [64, 79], [80, 95], [96, 111], [112, 127], [128, 143], [144, 159], [160, 175], [176, 191], [192, 207], [208, 223], [224, 239], and [240, 255]. The regional information module sets the image grayscale threshold Gp0 according to the number Y of intervals where each pixel point in the enhanced image is located, wherein When Y>4, the regional information statistics module sets the image grayscale threshold When Y≤4, the regional information statistics module sets the image grayscale threshold Gp0=0.5×(Gpmax+Gpmin); in, represents the average value of the first grayscale interval where the pixel point in the enhanced image is located, represents the average value of the second grayscale interval where the pixel point in the enhanced image is located, represents the average value of the i-th grayscale interval in which the pixel points in the enhanced image are located, c1 represents the number of pixels in the first grayscale interval, c2 represents the number of pixels in the second grayscale interval, ci represents the number of pixels in the i-th grayscale interval, Gpmax represents the maximum grayscale value of the pixel points in the enhanced image, and Gpmin represents the minimum grayscale value of the pixel points in the enhanced image.

8. The artificial intelligence-based object detection device according to claim 7, characterized in that: The regional information statistics module performs binarization processing on the enhanced image after grayscale processing to obtain the connected domain. The regional information module sets the image grayscale threshold to Gp0, and binarizes the grayscale of each pixel according to the comparison result of the grayscale value Gp of each pixel in the enhanced image after grayscale processing and the image grayscale threshold Gp0, wherein, When Gp≤Gp0, the regional information statistics module adjusts the grayscale of the corresponding pixel to 0; When Gp>Gp0, the region information statistics module adjusts the grayscale of the corresponding pixel to 255.

9. The artificial intelligence-based object detection device according to claim 8, characterized in that: The regional information statistics module selects the connected domain with the largest area Ymax, and obtains the area of ​​each connected domain and the pixel points located on the connected domain outline. The regional information statistics module obtains the minimum distance lk between the connected domain Ymax outline point and the connected domain outline, k=1,2,…n, n is the number of connected domains, and determines whether each connected domain is valid based on the area Sk of each connected domain and the minimum distance lk between the connected domain Ymax outline point and the connected domain outline, wherein, When Sk≤1 / 50×Smax, the region information statistics module determines that the connected domain is invalid and does not belong to the ROI region; When 1 / 50×Smax<Sk≤1 / 30×Smax and lk>dmax, the region information statistics module determines that the connected domain is invalid and does not belong to the ROI region; When 1 / 50×Smax<Sk≤1 / 30×Smax and lk≤dmax, the region information statistics module determines that the connected domain is valid and belongs to the ROI region; When Sk>1 / 30×Smax, the region information statistics module determines that the connected domain is valid and belongs to the ROI region; Among them, Smax represents the area of ​​the connected domain Ymax, and dmax represents the straight-line distance between the two pixels with the farthest straight-line distance on the contour of the connected domain Ymax.

10. The artificial intelligence-based object detection device according to claim 9, characterized in that: The regional information statistics module extracts each pixel point on the effective connected domain contour line, and selects the upper left vertex, the upper right vertex, the lower left vertex and the lower right vertex from all the pixel points on the effective connected domain contour line. The regional information statistics module connects the upper left vertex, the upper right vertex, the lower left vertex and the lower right vertex in a straight line so that each vertex is connected to form a quadrilateral. The regional information statistics module intercepts the ROI area according to the coverage area of ​​the quadrilateral, wherein, When the quadrilateral covers all valid connected domains, the regional information statistics module intercepts the quadrilateral as the ROI area; When the quadrilateral cannot cover all valid connected domains, the region information statistics module intercepts a bounding rectangle that can cover all valid connected domains as the ROI region.

Images