Hazardous article detection method based on millimeter wave human body image

Abstract

The invention discloses a hazardous article detection method based on millimeter wave human body images. The method comprises the steps that the millimeter wave human body images in a verification set are divided into different areas; independent threshold optimization is carried out on different areas divided by the millimeter wave human body image in the verification set, and a regional hazardous article detection threshold is constructed based on a threshold optimization result; performing preliminary dangerous article detection on the human body image according to a dangerous article detection model to obtain a dangerous article detection result, the detection result including the position, size, confidence and category of a dangerous article detection frame; for each dangerous goods detection frame, judging an area to which the detection frame belongs; for each detection frame, searching a threshold value corresponding to the category in the regional hazardous article detection threshold values according to the category of the detection frame; and filtering the detection result according to the confidence coefficient and a threshold value corresponding to the category, and outputting a final detection result. The hazardous article detection effect of the millimeter wave human body imaging equipment is improved.

Description

Technical Field

[0001] The present invention relates to the field of detecting dangerous goods on the human body surface by millimeter wave imaging, and more specifically, to a method for detecting dangerous goods based on millimeter wave human body images. Background Art

[0002] Millimeter waves are electromagnetic waves with a frequency band of 30GHz-300GHz. They can penetrate clothing fabrics, plastics and other materials, and can show different types of dangerous items such as knives, guns and explosives hidden in clothing, including metals, non-metals, solids and liquids. Millimeter wave images have good image contrast and spatial resolution, and can clearly show the location and shape of suspicious items. In addition, low doses of millimeter waves are harmless to human health. Therefore, millimeter wave-based human imaging is favored in the security inspection field and has been applied to various security inspection sites, such as airports, customs and large event venues.

[0003] Target detection algorithms are often used for dangerous goods detection using millimeter wave human body imaging. The main process is: reasoning based on millimeter wave human body images, outputting detection frames and confidences representing the location and size of dangerous goods; then threshold filtering based on confidence, retaining only detection frames with confidences not less than the corresponding threshold as the final dangerous goods detection results. Therefore, the confidence threshold plays a key filtering role and has a great impact on the detection rate and false alarm rate of dangerous goods. If the threshold is set too high, there will be fewer false alarms, but it is easy to cause missed detections; if the threshold is set too low, it is conducive to reducing missed detections, but it may cause more false alarms. Therefore, in the process of detecting dangerous goods based on millimeter wave human body imaging equipment, it is necessary to optimize the threshold of dangerous goods detection to obtain the best threshold. The existing threshold optimization method is based on the whole body area, and does not take into account the difference in imaging quality caused by the different degrees of microwave reflection of objects in different parts of the human body. In millimeter wave human body images, the reflectivity of objects on the forearm, upper arm, back of the head and calf of a person is lower than that of the chest and back, and the imaging quality is often relatively poor, making the detection of dangerous goods more difficult. Therefore, the differentiated characteristics of millimeter-wave imaging areas of the human body result in the existing threshold determination method based on the whole body area being unable to adapt well to the detection of dangerous goods in different areas of the human body. Summary of the invention

[0004] The present invention provides a dangerous goods detection method based on millimeter wave human body images to solve at least one of the problems existing in the prior art.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The present invention provides a method for detecting dangerous goods based on millimeter wave human body images, the method comprising:

[0007] Divide the millimeter-wave human images in the validation set into different regions;

[0008] Independently optimize the thresholds of different regions of the millimeter-wave human body images in the validation set, and build regional dangerous goods detection thresholds based on the results of threshold optimization;

[0009] Performing preliminary dangerous goods detection on the human body image output by the millimeter wave human body imaging device according to the dangerous goods detection model, obtaining detection results of dangerous goods detection frames, and for the detection results containing dangerous goods detection frames, determining the area to which each dangerous goods detection frame belongs, wherein the detection results include the position, size, confidence and category of the dangerous goods detection frame;

[0010] According to the category of the dangerous goods detection frame, searching for a threshold value corresponding to the category in the regional dangerous goods detection threshold dictionary;

[0011] The detection result is filtered according to the confidence and the threshold value corresponding to the category, and a final detection result is output.

[0012] Optionally, dividing the millimeter wave human body images in the verification set into different areas includes

[0013] In response to the acquired division data, the millimeter wave human body image is divided into regions; or

[0014] A human body key part detection model is trained according to the millimeter wave human body image and the annotations of the human body key parts, the key parts in the image are located based on the human body key part detection model, and different areas are divided according to the positioning.

[0015] Optionally, the independently optimizing the thresholds of different regions of the millimeter wave human image in the verification set includes:

[0016] Obtain the detection rate and false alarm rate based on the detection results of the dangerous goods labeling box and the dangerous goods detection model;

[0017] Constructing a threshold optimization objective function according to the detection rate and the false alarm rate;

[0018] The threshold optimization objective function is utilized, and a search method is selected to perform a threshold search in the region to obtain the best threshold in the region.

[0019] Optionally, the threshold optimization objective function f(τ1, ...τ N )for

[0020] f(τ1,...τ N )=wR(τ1,...τ N )-(1-w)F(τ1,...τ N ),

[0021] Where N is the number of categories, τ1,...τ N are the thresholds for dangerous goods categories 1 to N, R is the detection rate, F is the false alarm rate, and w is the balance coefficient between the detection rate and the false alarm rate.

[0022] Optionally, the construction area dangerous goods detection threshold includes

[0023] A regional dangerous goods detection threshold dictionary is constructed, different regions are numbered, and a regional dangerous goods detection threshold dictionary is constructed according to the region numbers and a set of thresholds of the corresponding regions.

[0024] Optionally, the detecting of the millimeter wave human image output by the millimeter wave human imaging device according to the dangerous goods detection model includes:

[0025] Obtaining detection results of the image to be detected according to the dangerous goods detection model;

[0026] The key parts of the image to be detected are located according to the human body key parts detection model to divide different areas.

[0027] Optionally, the region to which each dangerous goods detection frame belongs is determined to include

[0028] Calculate the overlapping area between the dangerous goods detection frame and different regions, take the region with the largest overlapping area as the region to which the dangerous goods detection frame belongs, obtain the region number to which the dangerous goods detection frame belongs, and obtain a set of thresholds for the dangerous goods detection frame in the regional dangerous goods detection threshold dictionary according to the region number.

[0029] Optionally, the area number to which the dangerous goods detection frame belongs is

[0030]

[0031] Among them, j is the region number to which the dangerous goods detection box belongs, k is the number of regions divided into the image, S i is the overlapping area between the dangerous goods detection frame and region i.

[0032] Optionally, searching the regional dangerous goods detection threshold dictionary for a threshold value corresponding to the category according to the category of the dangerous goods detection frame includes:

[0033] According to the category of the dangerous goods detection frame detection result, searching for a threshold value corresponding to the category in a set of threshold values ​​of the area to which the dangerous goods belong;

[0034] The detection result of the dangerous goods detection frame is filtered according to the confidence of the dangerous goods detection frame and the threshold.

[0035] Optionally, the detection result of screening the dangerous goods detection frame according to the confidence of the dangerous goods detection frame and the threshold includes:

[0036] Determine the size of the confidence of the dangerous goods detection frame and the threshold corresponding to the category of the dangerous goods detection frame, and if the confidence of the dangerous goods detection frame is greater than or equal to the threshold, retain the detection result of the dangerous goods detection frame;

[0037] If the confidence of the dangerous goods detection frame is less than the threshold, the detection result of the dangerous goods detection frame is discarded.

[0038] The beneficial effects of the present invention are as follows:

[0039] The present invention proposes a dangerous goods detection method based on millimeter wave human imaging equipment, which divides the millimeter wave human images in the verification set into regions with different characteristics, and then optimizes the threshold value independently for each region to obtain the optimal threshold value in each region, so that the dangerous goods detection threshold has regional adaptability; performs preliminary dangerous goods detection on the human image according to the dangerous goods detection model, obtains the dangerous goods detection result, determines the region to which each detection frame belongs, filters the detection result according to the confidence of the dangerous goods detection frame and the threshold value corresponding to the category of the detection frame, and outputs the final detection result, thereby improving the detection effect of the millimeter wave human imaging device in detecting dangerous goods. This method can also be applied to various target detection scenarios involving millimeter wave human images, and can also be expanded to other target detection application scenarios, such as industrial detection, medical diagnosis, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0041] Figure 1 A flow chart showing a dangerous goods detection method based on a millimeter wave human body imaging device according to the present invention;

[0042] Figure 2 A schematic diagram showing three regions for dividing a millimeter wave human body image in an embodiment of the present invention;

[0043] Figure 3 A schematic diagram showing millimeter wave human body images at different angles in an embodiment of the present invention;

[0044] Figure 4 A flow chart showing the present invention's dangerous goods detection for a millimeter wave human body image to be detected. DETAILED DESCRIPTION

[0045] In order to more clearly illustrate the present invention, the present invention is further described below in conjunction with preferred embodiments and accompanying drawings. Similar components in the accompanying drawings are represented by the same reference numerals. It should be understood by those skilled in the art that the content specifically described below is illustrative rather than restrictive, and should not be used to limit the scope of protection of the present invention.

[0046] Aiming at the problem that the threshold of dangerous goods detection in the existing millimeter wave human body imaging equipment cannot be adaptive to the region, the present invention proposes a dangerous goods detection method based on millimeter wave human body images, the method comprises:

[0047] Divide the millimeter-wave human images in the validation set into different areas to improve the detection accuracy of dangerous goods;

[0048] The method of dividing the area of ​​the millimeter wave human body image includes rough division and precise division, wherein the rough division is to divide the millimeter wave human body image into regions in response to the acquired division data, and the positions of the key parts of the front standing human body image of the millimeter wave human body image are estimated to divide the approximate regions, such as the head and arm region, the trunk region, the leg and foot region. This kind of division does not rely on the precise identification of the key parts of the human body. The advantages are simple operation and high efficiency, and the disadvantage is that the accuracy of the divided regions is limited. The precise division first trains the key parts detection model of the human body according to the millimeter wave human body image and the annotation of the key parts of the human body, and identifies and locates the key parts of the human body based on the trained key parts detection model, such as the identification of the head, hands, elbows, hips, knees and feet of the human body, and then divides the regions according to the needs. The advantages of this region division method are high accuracy of the divided regions and more guaranteed improvement effect. The region division frame for dividing the region can use different shapes, such as horizontal rectangles (the sides of the rectangle are parallel or perpendicular to the wide sides of the image), inclined rectangles (the sides of the rectangle are not parallel or perpendicular to the sides of the image), parallelograms and circles, etc.

[0049] Independently optimize the thresholds of different regions of the millimeter-wave human body images in the validation set, and build regional dangerous goods detection thresholds based on the results of threshold optimization;

[0050] Obtain the detection rate and false alarm rate based on the detection results of the dangerous goods labeling box and the dangerous goods detection model;

[0051] Construct a threshold optimization objective function based on the detection rate and false alarm rate;

[0052] By using the threshold optimization objective function, an appropriate search method is selected to perform threshold search in the region to obtain the best threshold in the region.

[0053] For each area, based on the known data set, according to the annotation content of the dangerous goods annotation box and the dangerous goods detection model, the hyper-parameter optimization method is used to independently search for the optimal dangerous goods detection threshold in the area. When the detection threshold optimization on all areas is completed, different areas are numbered, and a regional dangerous goods detection threshold dictionary is constructed according to the area number and the threshold of the corresponding area, so as to find the optimal detection threshold for different areas.

[0054] A threshold optimization objective function is constructed according to the detection rate and the false alarm rate, and threshold optimization is performed on different areas according to the threshold optimization objective function to obtain the best detection thresholds for different areas.

[0055] Threshold optimization objective function f(τ1,...τ N )for

[0056] f(τ1,...τ N )=wR(τ1,...τ N )-(1-w)F(τ1,...τ N ),

[0057] Where N is the number of categories, τ1,...τ N are the thresholds for dangerous goods categories 1 to N, R is the detection rate, F is the false alarm rate, and w is the balance coefficient between the detection rate and the false alarm rate.

[0058] According to the hazardous goods detection model, a preliminary hazardous goods detection is performed on the human body image output by the millimeter-wave human imaging device. If the preliminary detection result is empty, that is, the millimeter-wave human body image does not contain a hazardous goods detection frame, the detection is terminated; if the preliminary detection result is not empty, the detection results of all hazardous goods detection frames are obtained, and the detection results include the position, size, confidence and category of the hazardous goods detection frame, and the area to which each hazardous goods detection frame belongs is determined.

[0059] According to the dangerous goods detection model, the position, size, confidence level and category of the dangerous goods detection frame in the millimeter wave human image to be detected are obtained;

[0060] The key parts of the image to be detected are located according to the key parts detection model of the human body to divide different areas;

[0061] Calculate the overlapping area between the dangerous goods detection frame and different regions, take the region with the largest overlapping area as the region to which the dangerous goods detection frame belongs, obtain the region number to which the dangerous goods detection frame belongs, obtain a set of thresholds for the dangerous goods detection frame in the regional dangerous goods detection threshold dictionary according to the region number, and search for the threshold corresponding to the category in the set of thresholds for the dangerous goods detection frame according to the category in the detection result of the dangerous goods detection frame.

[0062] Get the area number to which the dangerous goods detection frame belongs

[0063]

[0064] Among them, j is the region number to which the dangerous goods detection box belongs, k is the number of regions divided into the image, S i is the overlapping area between the dangerous goods detection frame and region i.

[0065] Filtering the detection result according to the confidence of the dangerous goods detection frame and the threshold value corresponding to the category of the dangerous goods detection frame, and outputting the final detection result;

[0066] The detection results of the dangerous goods detection frame are screened according to the confidence of the dangerous goods detection frame and the threshold, and the size of the confidence of the dangerous goods detection frame and the threshold corresponding to the category of the dangerous goods detection frame is determined. If the confidence of the dangerous goods detection frame is greater than or equal to the threshold, the detection result of the dangerous goods detection frame is retained; if the confidence of the dangerous goods detection frame is less than the threshold, the detection result of the dangerous goods detection frame is discarded. All the detection results that meet the conditions are used as the final detection results.

[0067] In a specific embodiment, Figure 1 As shown in the figure, firstly, the millimeter wave human body images in the verification set are divided into different regions according to the division method selected as needed; the threshold of dangerous goods detection is optimized independently for each region to obtain the optimal threshold of the corresponding region; then the threshold of dangerous goods detection in the entire image is constructed based on the threshold of each region. Among them, the dangerous goods to be detected can be determined according to the specific application requirements, such as flammable and explosive items, knives, guns, drugs, etc., and the detection category can be one or more.

[0068] The millimeter wave human imaging device in this embodiment is a cylindrical millimeter wave human body scanning system. The cylindrical millimeter wave human body scanning system scans and images the human body at multiple angles. In addition to the front and back images, there are also images at different angles such as the side front and side back. Dangerous goods can be detected at multiple angles, thereby reducing missed detections and false alarms. The dangerous goods detection model in this embodiment can be a deep learning model, such as the Fast RCNN series, the YOLO series detection model, or a traditional detection model, such as a sliding window-based detection model.

[0069] The millimeter-wave human images in the validation set are divided into different areas. You can choose between a rough division based on manual estimation of human body parts and a precise division based on algorithmic recognition of human body parts. When dividing the millimeter-wave human body images, the areas are divided according to the image quality and recognition difficulty of dangerous goods in different areas. Since the image quality of items in the forearms, upper arms, back of the head and calves is generally poor, this embodiment adopts a precise division method to divide the millimeter-wave human body images output by the millimeter-wave human body imaging device into regions. Specifically, the millimeter-wave human body images can be divided into three areas, namely the head and arm area, the torso and thigh area, and the calf area. Among them, the head and arm area includes the head, forearms, upper arms and armpits; the torso and thigh area includes the human torso, buttocks and thighs; the calf area includes the calf and foot. As Figure 2The figure shows the three regions divided by the millimeter wave human body image, where the face and crotch are covered with a mosaic pattern to protect privacy, and the head and arm region, the trunk and thigh region, and the calf region are recorded as region 1, region 2, and region 3, respectively.

[0070] Assume that the width of the image is w, the height is h, the upper left corner of the image is the origin, the horizontal right direction is the positive direction of the x-axis, and the vertical downward direction is the positive direction of the y-axis. The entire rectangular image is recorded as (x min ,y min ,x max ,y max ), where x min and min Respectively represent the x-axis and y-axis coordinates of the upper left vertex of the rectangular image, x max and max Respectively represent the x-axis and y-axis coordinates of the lower right vertex of the rectangular image. Let the height of the head and arm region be h1, then the head and arm region is represented as (0, 0, w, h1), let the height of the calf region be h2, then the calf region can be represented as (0, h-h2, w, h), and the rectangular region corresponding to the trunk and thigh region is represented as (0, h1, w, h-h2).

[0071] Rough division estimates the height h1 of the head and arm area and the height h2 of the calf area based on experience. When using rough division to divide the millimeter wave human body image into regions, all images use the same parameters h1 and h2. The height h1 of the head and arm area can be taken as h / 3, and the height h2 of the calf area can be taken as 3h / 4. The accuracy of the regions obtained by rough division is limited. Accurate division trains a human key part detection model based on millimeter wave human images and annotations of key parts of the human body, locates the key parts in the image based on the human key part detection model, and divides different regions according to the positioning.

[0072] According to the millimeter wave human body image and the annotation of the key parts of the human body, a model that can identify the key parts of the human body in the millimeter wave image is trained, that is, the key parts detection model of the human body. The trained key parts detection model of the human body is used to identify and locate the key parts of the human body in the input image, and a more accurate area is divided. Specifically, the key parts in the millimeter wave human body image are annotated, such as the head, shoulders, knees and other parts; according to the millimeter wave human body image and the annotation of the key parts, a depth detection model such as YOLOv8s is selected to train the key parts detection model of the human body; the height h1 of the head and arm area is estimated by locating the left and right shoulders, and the height h2 of the calf area is estimated by locating the knees. The ordinates of the lower right vertices of the left and right shoulder detection frames are y1 and y2 respectively, then the height h1 of the head and arm area is max(y1,y2), and the ordinates of the upper left vertices of the left and right knee detection frames are y3 and y4 respectively, then the height h2 of the calf area is min(y3,y4).

[0073] like Figure 3 As shown, Figure 3 (a) shows the three regions of the front of the millimeter wave human body image. Figure 3 (b) is the area divided at different angles of the side or back. For millimeter wave human body images at different angles, based on Figure 3 (a) The three regions of the front image are used as reference. The face and crotch are covered with mosaic patterns to protect privacy. The image is divided into three regions based on the different reflectivity of different parts of the human body to objects. This is only an example of region division. Different region divisions can also be performed based on the characteristics and requirements of the device's millimeter wave human image.

[0074] The threshold value of dangerous goods detection is optimized independently for different divided areas. For each area, the detection results of the dangerous goods labeling box and the dangerous goods detection model that are concentrated in the area are used to obtain the detection rate and false alarm rate of the area. The threshold optimization objective function is constructed based on the searched detection rate and false alarm rate. Based on the threshold optimization objective function, an appropriate search method is selected to optimize the thresholds of different areas independently to obtain the best detection threshold for the area. The threshold optimization objective function f(τ1,...τ N )for

[0075] f(τ1,...τ N )=wR(τ1,...τ N )-(1-w)F(τ1,...τ N ),

[0076] Where N is the number of categories, τ1,...τ N are the thresholds for dangerous goods categories 1 to N, R is the detection rate, F is the false alarm rate, and w is the balance coefficient between the detection rate and the false alarm rate.

[0077] The search method for threshold search can be grid search, coordinate descent, Bayesian optimization, etc.

[0078] Constructing a dictionary of threshold values ​​for dangerous goods detection is only one way of managing threshold values, and threshold values ​​can also be managed in other ways. This embodiment adopts the method of constructing a dictionary of threshold values ​​for dangerous goods detection, numbers different regions, and constructs a regional dictionary of dangerous goods detection threshold values ​​using the region number and the best detection threshold value obtained for the corresponding region. The key value in the dictionary is the region number, and the value is a set of threshold values ​​for the corresponding region, so that the subsequent dangerous goods detection algorithm can search for threshold values ​​according to the region.

[0079] like Figure 4 As shown, a preliminary dangerous goods detection is performed on the human body image output by the millimeter wave human body imaging device according to the dangerous goods detection model. If the preliminary detection result is empty, that is, no dangerous goods detection frame is detected in the millimeter wave human body image, the detection is terminated; if the preliminary detection result is not empty, the detection results of all dangerous goods detection frames are obtained, wherein the detection results of the dangerous goods detection frame include the position, size, confidence and category of the dangerous goods detection frame, and the area to which each dangerous goods detection frame belongs is determined;

[0080] According to the dangerous goods detection model, the input human body image is identified and located to obtain the detection results of the dangerous goods detection model; according to the human body key part detection model, the key parts of the image to be detected are located, so as to divide the image to be detected into the head and arm area, the torso and thigh area, and the calf area. Calculate the overlapping area S of the dangerous goods detection frame and the three divided areas i , the largest area in the overlapping area is taken as the area to which the dangerous goods detection frame belongs. In this embodiment, the image is divided into three areas, that is, k = 3, and the area numbers are recorded as 1, 2, and 3 respectively. The overlapping areas of a certain dangerous goods detection frame and the three areas are recorded as S1, S2, and S3 respectively. The area number to which the dangerous goods detection frame belongs is

[0081]

[0082] Among them, j is the region number to which the dangerous goods detection box belongs, k is the number of regions divided into the image, S i is the overlapping area between the dangerous goods detection frame and region i.

[0083] Obtain a set of thresholds for the dangerous goods detection frame in the regional dangerous goods detection threshold dictionary according to the regional number, and search for the threshold corresponding to the category in a set of thresholds for the region to which the dangerous goods belong according to the category information in the detection result;

[0084] Determine the size of the confidence in the detection result of the dangerous goods detection frame and the detection threshold corresponding to the area to which the dangerous goods detection frame belongs. If the confidence of the dangerous goods detection frame is greater than or equal to the detection threshold, retain the dangerous goods detection frame and output the detection result; if the confidence of the dangerous goods detection frame is less than the detection threshold, discard the dangerous goods detection frame.

[0085] Specifically, for a certain detection result, a set of thresholds for the area to which the dangerous goods detection box belongs is τ1,...τ N , corresponding to categories 1 to N respectively, find the threshold τ corresponding to category i in the detection result of the dangerous goods detection frame i (1≤i≤N), determine the confidence of the dangerous goods detection frame and the corresponding threshold τ of the category of the dangerous goods detection frame i If the confidence of the dangerous goods detection frame is greater than or equal to the threshold, the detection result of the dangerous goods detection frame is retained; if the confidence of the dangerous goods detection frame is less than the threshold, the detection result of the dangerous goods detection frame is discarded. All the detection results that meet the conditions are taken as the final detection results.

[0086] The present invention proposes a dangerous goods detection method based on millimeter wave human body images. By dividing the millimeter wave human body images into regions with different characteristics, each region is then independently optimized to obtain a threshold value that is more suitable for the region, so that the dangerous goods detection threshold has regional adaptability, thereby improving the detection rate of dangerous goods detection and reducing the false detection rate. The method can also be applied to various target detection scenarios involving millimeter wave human body imaging equipment, such as security inspection, crowd monitoring, baggage inspection, etc. The detection threshold regional adaptation method can also be extended to other target detection application scenarios, such as industrial inspection, medical diagnosis, etc.

[0087] In the description of the present invention, it should be noted that the terms "upper", "lower", etc. indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific position, be constructed and operated in a specific position, and therefore cannot be understood as a limitation on the present invention. Unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be a connection between the two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0088] It should also be noted that, in the description of the present invention, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0089] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not limitations on the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the protection scope of the present invention.

Claims

1. A method for detecting dangerous goods based on millimeter wave human body images, characterized in that: The method includes Divide the millimeter-wave human images in the validation set into different regions; Independently optimize the thresholds of different regions of the millimeter-wave human body images in the validation set, and build regional dangerous goods detection thresholds based on the results of threshold optimization; Perform preliminary dangerous goods detection on the millimeter wave human image output by the millimeter wave human imaging device according to the dangerous goods detection model, obtain the detection result of the dangerous goods detection frame, the detection result includes the position, size, confidence and category of the dangerous goods detection frame, and determine the area to which each dangerous goods detection frame belongs; According to the category of the dangerous goods detection frame, a threshold value corresponding to the category is searched in the regional dangerous goods detection threshold dictionary, the detection result is filtered according to the confidence and the threshold value corresponding to the category, and a final detection result is output.

2. The method according to claim 1, characterized in that The method of dividing the millimeter wave human body images in the verification set into different areas includes: In response to the acquired division data, the millimeter wave human body image is divided into regions; or A human body key part detection model is trained according to the millimeter wave human body image and the annotations of the human body key parts, the key parts in the image are located based on the human body key part detection model, and different areas are divided according to the positioning.

3. The method according to claim 1, characterized in that The independent threshold optimization of different regions of the millimeter wave human image in the validation set includes Obtain the detection rate and false alarm rate based on the detection results of the dangerous goods labeling box and the dangerous goods detection model; Constructing a threshold optimization objective function according to the detection rate and the false alarm rate; The threshold optimization objective function is utilized and a search method is selected to perform a threshold search in the region to obtain the best threshold in the region.

4. The method according to claim 3, characterized in that The threshold optimization objective function f(τ1,...τ N )for f(τ1,...τ N )=wR(τ1,...τ N )-(1-w)F(τ1,...τ N ), Where N is the number of categories, τ1,...τ N are the thresholds for dangerous goods categories 1 to N, R is the detection rate, F is the false alarm rate, and w is the balance coefficient between the detection rate and the false alarm rate.

5. The method according to claim 1, characterized in that The construction area dangerous goods detection threshold includes A regional dangerous goods detection threshold dictionary is constructed, different regions are numbered, and a regional dangerous goods detection threshold dictionary is constructed according to the region numbers and a set of thresholds of the corresponding regions.

6. The method according to claim 1, characterized in that The method of performing preliminary dangerous goods detection on the millimeter wave human body image output by the millimeter wave human body imaging device according to the dangerous goods detection model includes: Obtaining detection results of the image to be detected according to the dangerous goods detection model; The key parts of the image to be detected are located according to the human body key parts detection model to divide different areas.

7. The method according to claim 1, characterized in that The area to which each dangerous goods detection frame belongs includes: Calculate the overlapping area between the dangerous goods detection frame and different regions, take the region with the largest overlapping area as the region to which the dangerous goods detection frame belongs, obtain the region number to which the dangerous goods detection frame belongs, and obtain a set of thresholds for the dangerous goods detection frame in the regional dangerous goods detection threshold dictionary according to the region number.

8. The method according to claim 7, characterized in that The area number to which the dangerous goods detection frame belongs is obtained as follows: Among them, j is the region number to which the dangerous goods detection box belongs, k is the number of regions divided into the image, S i is the overlapping area between the dangerous goods detection frame and region i.

9. The method according to claim 1, characterized in that: The searching of the threshold value corresponding to the category in the regional dangerous goods detection threshold dictionary according to the category of the dangerous goods detection frame includes: According to the category of the dangerous goods detection frame detection result, searching for a threshold value corresponding to the category in a set of threshold values ​​of the area to which the dangerous goods belong; The detection result of the dangerous goods detection frame is filtered according to the confidence of the dangerous goods detection frame and the threshold.

10. The method according to claim 9, characterized in that The detection result of the dangerous goods detection frame is screened according to the confidence of the dangerous goods detection frame and the threshold value, including Determine the size of the confidence of the dangerous goods detection frame and the threshold corresponding to the category of the dangerous goods detection frame, and if the confidence of the dangerous goods detection frame is greater than or equal to the threshold, retain the detection result of the dangerous goods detection frame; If the confidence of the dangerous goods detection frame is less than the threshold, the detection result of the dangerous goods detection frame is discarded.

Images