Human body security check method and system

Abstract

The invention discloses a human body security check method and system, and the method comprises the steps: constructing an abnormal image detection model, and training the abnormal image detection model based on a first training data set, so as to obtain a trained abnormal image detection model; acquiring original image data in the scanning area; inputting the original image data into a trained abnormal image detection model to judge whether the original image data is human body data or not, and taking the human body data as an abnormal image judgment result; and outputting an abnormal image discrimination result to the outside.

Description

Technical Field

[0001] The present invention relates to a human body security inspection method and system, and more specifically, to a human body security inspection method and system based on deep learning and training of a model. Background Art

[0002] Whether the passenger's posture during security check is standard is related to the efficiency of security check. Using security inspectors to supervise passenger security check will increase the communication cost between passengers and security inspectors, increase the workload of security inspectors, and have the disadvantages of low security check pass rate.

[0003] Most existing image recognition methods use traditional image recognition processing methods, that is, extracting features from wrong posture images or abnormal images through manual feature extraction, and classifying them through classifiers. However, the accuracy of feature selection will directly affect the recognition accuracy. If too many features are selected, the dimension of the feature vector will increase, and the difficulty and complexity of the calculation will increase accordingly. On the contrary, if too few features are selected, the recognition accuracy may not be high. In addition, manual feature extraction is labor-intensive and complex, and has high requirements for application scenarios, and is of poor practicality. Summary of the invention

[0004] The purpose of the present invention is to provide a human body security inspection method and system. The method can utilize a deep learning convolutional neural network structure model, thereby having improved error posture discrimination accuracy, fast algorithm detection speed, and high practicality.

[0005] According to one aspect of the present invention, a human body security inspection method is provided, the method comprising: constructing an abnormal image detection model, and training the abnormal image detection model based on a first training data set to obtain a trained abnormal image detection model; acquiring original image data within a scanning area; inputting the original image data into the trained abnormal image detection model to determine whether the original image data is human body data as an abnormal image discrimination result; and outputting the abnormal image discrimination result to the outside.

[0006] In an embodiment, training the abnormal image detection model based on the first training data set includes: collecting multiple human data and multiple non-human data, and adding attribute category labels to each collected data to form the first training data set; inputting the first training data set into the abnormal image detection model, so that the abnormal image detection model judges each data in the first training data set, and trains with reference to the corresponding attribute category labels.

[0007] In an embodiment, the plurality of human body data includes human body data with a correct standing posture and human body data with an incorrect standing posture, and the plurality of non-human body data includes a plurality of different types of object data.

[0008] In an embodiment, determining whether the original image data is human body data includes: calculating the overall confidence of the original image data and comparing the overall confidence with a first threshold; and when the overall confidence is greater than the first threshold, determining that the original image data is human body data.

[0009] In an embodiment, a standing posture detection model is constructed, and the standing posture detection model is trained based on a second training data set to obtain a trained standing posture detection model; when it is determined that the original image data is the human body data, the original image data is input into the trained standing posture detection model to determine whether the human body standing posture is correct as a standing posture discrimination result; and the standing posture discrimination result is output to the outside.

[0010] In an embodiment, training the standing posture detection model based on a second training data set includes: collecting a plurality of standing posture image data, wherein the plurality of standing posture image data include standing posture image data of a plurality of incorrect standing posture types, and each incorrect standing posture type is assigned a standing posture type label; performing image preprocessing on the plurality of standing posture image data to form the second training data set; and inputting the second training data set into the standing posture detection model, so that the standing posture detection model judges each standing posture image data in the second training data set, and trains the standing posture detection model with reference to the corresponding standing posture type label.

[0011] In an embodiment, image preprocessing of the multiple standing posture image data includes at least one of the following: performing screening processing on the multiple standing posture image data collected to delete invalid standing posture image data; performing homogenization processing on the multiple standing posture image data collected to unify the number of standing posture image data of each erroneous standing posture type; performing expansion processing on the multiple standing posture image data collected; performing normalization processing on the multiple standing posture image data collected; and performing randomization processing on the multiple standing posture image data collected.

[0012] In an embodiment, determining whether a human body posture is correct includes: calculating the confidence level for each of the multiple incorrect posture types according to the original image data; when all the calculated confidence levels are lower than a second threshold, determining that the human body posture is correct; when all the calculated confidence levels are not lower than the second threshold, determining that the human body posture is incorrect, and taking the incorrect posture type with the highest confidence level as the posture determination result.

[0013] In an embodiment, when it is determined that the original image data is not the human body data, new original image data is recaptured, and the new original image data is re-input into the trained abnormal image detection model to determine whether it is human body data.

[0014] In an embodiment, the method further comprises: outputting the abnormal image identification result and / or the standing posture identification result to the outside in a visual display manner or a voice prompt manner, so as to provide an error prompt during a human body security inspection.

[0015] In an embodiment, the abnormal image detection model and the standing posture detection model are Resnet classification network models.

[0016] According to another aspect of the present invention, a human body security inspection system is provided, which executes the human body security inspection method according to an embodiment of the present invention, and the system includes: an image acquisition device, which is configured to obtain original image data in a scanning area; and a computing device, which is configured to train an abnormal image detection model to obtain a trained abnormal image detection model, and to enable the trained abnormal image detection model to perform calculations to obtain abnormal image discrimination results.

[0017] In an embodiment, the computing device is further configured to train the stance detection model to obtain a trained stance detection model, and to enable the trained stance detection model to perform calculations to obtain a stance discrimination result.

[0018] In an embodiment, the system also includes: a visual display device, which is configured to output the abnormal image discrimination result and / or the standing posture discrimination result to the outside in a visual display manner; and / or a voice prompt device, which is configured to output the abnormal image discrimination result and / or the standing posture discrimination result to the outside in a voice prompt manner.

[0019] In an embodiment, the image acquisition device is also configured to acquire multiple standing posture image data; and the computing device also includes a data processing unit, which is configured to perform image preprocessing on the multiple standing posture image data to form a second training data set for training the standing posture detection model.

[0020] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings are used to better understand the present invention and do not constitute a limitation of the present invention, wherein:

[0022] Figure 1 A flow chart of a human body security inspection method according to an embodiment of the present invention is shown.

[0023] Figure 2 A training flowchart of an abnormal image detection model according to an embodiment of the present invention is shown.

[0024] Figure 3 A training flow chart of a standing posture detection model according to an embodiment of the present invention is shown.

[0025] Figure 4 An example of a residual network structure is shown.

[0026] Figure 5A and Figure 5B Schematic diagrams of two residual network structures combined according to embodiments of the present invention are respectively shown.

[0027] Figure 6 A schematic diagram showing the network structure of a Resnet classification network model according to an embodiment of the present invention is shown.

[0028] Figure 7 A schematic diagram of a human body security inspection system according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0029] In order to more clearly explain the purpose, technical solutions and advantages of the present invention, the embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings. It should be understood that the following description of the embodiments is intended to explain and illustrate the overall concept of the present invention, and should not be understood as limiting the present invention. In the specification and the drawings, the same or similar reference numerals refer to the same or similar parts or components. For the sake of clarity, the drawings are not necessarily drawn to scale, and some well-known parts and structures may be omitted in the drawings.

[0030] Unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. The wording "one" or "an" does not exclude multiple. "Including" or "comprising" and similar words mean that the elements or objects appearing in front of the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up", "down", "left", "right", "top" or "bottom" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly. When an element such as a layer, a film, a region or a substrate substrate is referred to as being "on" or "under" another element, the element may be "directly" located "on" or "under" another element, or there may be an intermediate element.

[0031] Figure 1A flow chart of a human body security inspection method according to an embodiment of the present invention is shown.

[0032] In step S1, an abnormal image detection model IDM can be constructed, and the abnormal image detection model IDM can be trained based on the first training data set DA1 to obtain a trained abnormal image detection model IDM'. In this article, the "abnormal image detection model IDM" can be understood as a calculation model for determining whether the image data collected by the image acquisition device in the scanning area is human body data. For example, in some cases, the detected person may stand in the wrong position during the security check and deviate from the scanning area, resulting in the failure to collect the required human body data. In some other cases, the detected person may also deviate from the scanning area due to reasons such as height. This will make it impossible to continue the security check process, so it is necessary to first ensure that the detected person is in the correct standing position, that is, within the scanning area of ​​the image acquisition device. According to an embodiment of the present invention, the detected person can be ensured to be in the correct standing position by using the abnormal image detection model IDM to determine whether the collected original image data OIM is human body data. Therefore, when the detected person correctly stands in or is in the scanning area of ​​the image acquisition device, the collected original image data OIM is the human body data of the detected person, rather than the nearby object data (that is, non-human body data).

[0033] Figure 2 A training flowchart of an abnormal image detection model IDM according to an embodiment of the present invention is shown.

[0034] like Figure 2 As shown, in step S121, multiple human body data and multiple non-human body data can be collected, and attribute category labels are added to each collected data to form a first training data set DA1. As an example, various types and quantities of objects can be arranged in a human body security inspection scene (for example, a security inspection channel) or in a similar scene, and can be arranged in various ways and positions so that the image acquisition device can collect multiple non-human body data. For example, multiple different types of objects included in the multiple non-human body data include: cartons, chairs, stools, standard test pieces, test stands, trash cans, chassis, isolation belts, and the like. However, embodiments of the present invention are not limited to this.

[0035] On the other hand, the image acquisition device can also be used to collect multiple human body data in the scanning area. As an example, the multiple human body data can include human body data with correct standing posture and human body data with incorrect standing posture. In some cases, although the detected person may be incorrect in standing posture, it can still be detected as human body data when the standing position is correct, so that feedback that the position of the detected person is correct can be output to the outside (i.e., the detection person).

[0036] Afterwards, attribute category labels may be added to each collected image data to identify whether each image data is a person category or an object category. As an example, each collected image data and its respective attribute category label may be formed into a single entry, and then all entries may be combined into the first training data set DA1.

[0037] Generally, the newly constructed abnormal image detection model IDM may not be reliable in its recognition ability of the image data collected in the human body security inspection, so it needs to be trained with relevant types of data to have accurate and reliable recognition ability. Therefore, according to an embodiment of the present invention, by using the image acquisition device (for example, a millimeter wave human body security inspection instrument) in the human body security inspection instrument to pre-collect a plurality of original image data, and pre-identify the recognition results (i.e., attribute category labels) for each collected data, and then by using these data to train the abnormal image detection model IDM, the trained abnormal image detection model IDM' can have accurate and reliable recognition ability for the human body security inspection image data, thereby improving the ability to prompt the wrong standing position in the human body security inspection.

[0038] In step S122, the first training data set DA1 can be input into the abnormal image detection model IDM, so that the abnormal image detection model IDM judges each data in the first training data set DA1 and trains with reference to the corresponding attribute category label. As an example, the abnormal image detection model IDM can first make a type judgment (for example, judge it as human body data) on the first image data in the first training data set DA1 (for example, the first entry in the first training data set DA1), and calculate the difference value (i.e., loss value) between the judgment result and the corresponding attribute category label of the first image data. The parameters of the abnormal image detection model IDM are continuously updated iteratively by minimizing the difference value to form an updated abnormal image detection model IDM, and the updated abnormal image detection model IDM is used to recycle until the loss value meets a certain minimum condition (for example, the loss value is reduced to below a certain threshold, or the number of pre-training is reached). By analogy, the cycle is completed for each data in the first training data set DA1, thereby completing the training process of the abnormal image detection model IDM. The abnormal image detection model IDM obtained when all the cycles of the first training data set DA1 are completed is the trained abnormal image detection model IDM'.

[0039] Next, return Figure 1In step S2, the original image data OIM in the scanning area can be obtained. As an example, the image acquisition device in the human body security inspection equipment can be used to scan the predetermined area to directly acquire the original image data OIM. For example, the millimeter wave human body security inspection device in the millimeter wave human body security inspection equipment can be used for scanning. According to an embodiment of the present invention, the acquired original image data can be directly used in the subsequent model calculation process without the need for other data processing.

[0040] In step S3, the original image data OIM can be input into the trained abnormal image detection model IDM' to determine whether the original image data OIM is human data as the abnormal image discrimination result IMR. As an example, as described above, the trained abnormal image detection model IDM' already has sufficient ability to accurately identify the type of the original image data OIM. Therefore, in the human body security inspection process, after the image acquisition device acquires the original image data OIM in the scanning area, the original image data OIM is directly input into the trained abnormal image detection model IDM' to determine whether the original image data OIM is human data, and generate the corresponding abnormal image discrimination result IMR. For example, the abnormal image discrimination result IMR can prompt the detection personnel that the original image data OIM collected at this time is human data, that is, the current position of the detected person is correct, and the subsequent detection process can be continued. Alternatively, the abnormal image discrimination result IMR can prompt the detection personnel that the original image data OIM collected at this time is non-human data, that is, the current position of the detected person is wrong, and the image acquisition of the current detected person needs to be performed again.

[0041] In one embodiment, after the original image data OIM is input into the trained abnormal image detection model IDM', the overall confidence of the original image data OIM can be calculated to make a type judgment according to the calculated overall confidence. For example, the trained abnormal image detection model IDM' can perform feature extraction processing on the original image data OIM, and perform score statistics on the extracted features to use the calculated score as the overall confidence. However, the embodiments of the present invention are not limited thereto, and the overall confidence of the original image data OIM can be calculated in any suitable manner. Then, the calculated overall confidence can be compared with the first threshold value to determine whether the original image data OIM is human body data or non-human body data according to the comparison result. As an example, the score value of the human body-related feature can be set to be relatively high, while the score value of the item-related feature can be set to be relatively low. In this case, when the original image data OIM is human body data, the calculated overall confidence will be relatively high, that is, it can be higher than the first threshold; and when the original image data OIM is non-human body data, the calculated overall confidence will be relatively low, that is, it can be lower than the first threshold. In this way, the type judgment of the original image data OIM can be achieved.

[0042] According to an embodiment of the present invention, step S1 may further include constructing a standing posture detection model POM, and training the standing posture detection model POM based on a second training data set DA2 to obtain a trained standing posture detection model POM'. In this article, the "standing posture detection model POM" can be understood as an algorithm model used to determine whether the standing posture of the detected person meets the security inspection standards. For example, in some cases, the security inspection equipment will perform a full-body scan on the detected person to check whether contraband is carried. This usually requires the detected person to stand in a prescribed standard action to ensure that the security inspection equipment can scan all necessary parts of the human body to ensure the accuracy and pass rate of the security inspection. Therefore, the standing posture detection model POM can determine whether the standing posture of the detected person meets the requirements before the security inspection scanning process, and if it is prompted that it does not meet the requirements, the detected person can adjust himself or the detection personnel can assist in adjusting to the correct posture, and then perform a security inspection scan. In this way, the efficiency, accuracy and reliability of security inspection can be effectively improved.

[0043] Figure 3 A training flowchart of the standing posture detection model POM according to an embodiment of the present invention is shown.

[0044] like Figure 3As shown, in step S131, multiple standing posture image data can be collected. The multiple standing posture image data can include standing posture image data of multiple wrong standing posture types, and each wrong standing posture type is assigned a standing posture type label. As an example, the image acquisition device (for example, millimeter wave human body security detector) in human body security inspection can still be used, and the acquisition personnel can enter the scanning area and make multiple standing postures in sequence to collect image data. For example, the acquisition personnel can make multiple types of wrong standing postures, such as left hand error, right hand error, both hands out of bounds, both hands and arms are too close to the body, both hands are too close to the body, leg posture error, etc. After collecting the standing posture image data of these wrong standing posture types, each standing posture image data can be assigned a standing posture type label to identify the corresponding error type. For example, the standing posture error type can be represented by a pre-set letter or number, so as to be used as a standing posture type label. In this way, when it is identified that the detected person has a standing posture error, the specific error part (that is, the error type) can also be determined, so as to quickly correct and adjust the standing posture of the detected person in a targeted manner to smoothly advance the security inspection process.

[0045] In addition, the collector can also make a correct standing posture to collect one or more standing posture image data in the correct standing posture to increase the training dimension of the standing posture detection model POM. In addition, it should be noted that the embodiments of the present invention are not limited to the wrong standing posture types listed above, and other wrong standing posture types can be added according to specific actual needs to expand the recognition scope.

[0046] In step S132, the plurality of standing posture image data may be subjected to image preprocessing to form a second training data set DA2. In this article, "image preprocessing" may be understood as processing the plurality of collected standing posture image data in terms of training quality. That is, by processing the plurality of standing posture image data in terms of quantity and / or format, the processed plurality of standing posture image data may provide better training samples for the standing posture detection model POM, so that the trained standing posture detection model POM' has improved recognition performance. As an example, after performing such image preprocessing on the plurality of standing posture image data, updated plurality of standing posture image data may be obtained. These updated plurality of standing posture image data and their respective corresponding standing posture type labels are respectively formed into single entries, and then all entries are combined into the second training data set DA2.

[0047] In one embodiment, the image preprocessing of the plurality of standing posture image data may include at least one of the following: 1) screening processing, wherein the plurality of standing posture image data may be screened according to a predetermined standard to delete invalid standing posture image data. For example, standing posture image data with poor acquisition quality that cannot be used in subsequent calculation processes or redundant standing posture image data may be deleted to improve training efficiency. 2) homogenization processing, wherein the number of standing posture image data of each wrong standing posture type may be unified so that the proportion of each wrong standing posture type is substantially equal to improve the balance of training. Additionally, the number of standing posture image data of the correct standing posture may also be unified with the number of other types. 3) expansion processing, wherein the standing posture image data may be randomly flipped, rotated, scaled, etc. to expand the amount of training data. For example, expansion processing may be performed for each wrong standing posture type respectively so that the training dimension of each wrong standing posture type is maximized, thereby enabling accurate recognition of standing posture image data presented in various forms. Additionally, the above-mentioned homogenization processing may also be achieved by expansion. That is, by expanding each type, the number of each type is unified. 4) Normalization processing can make each standing posture image data in a predetermined standard format, that is, standardize each standing posture image data. For example, the RGB range of each standing posture image data can be standardized to be within the RGB range of 0~1. In this way, the processing complexity of the standing posture detection model POM for the standing posture image data can be reduced, thereby speeding up the calculation speed and training process. 5) Randomization processing can randomly shuffle multiple standing posture image data to enhance the training intensity and improve the recognition performance of the trained standing posture detection model POM'.

[0048] Similar to the above-mentioned abnormal image detection model IDM, the newly constructed standing posture detection model POM may not be reliable in its recognition ability of the image data collected in the human body security inspection, so it needs to be trained with relevant types of data to have accurate and reliable recognition capabilities. Therefore, according to an embodiment of the present invention, by using the image acquisition device in the human body security inspection instrument (for example, a millimeter wave human body security inspection instrument), a plurality of standing posture image data are pre-collected, and the recognition results (i.e., standing posture type labels) are pre-labeled for each data. Then, by using these data to train the standing posture detection model POM, the trained standing posture detection model POM' can have accurate and reliable recognition capabilities for human body security inspection image data, so as to determine whether the standing posture of the detected person is correct, and to prompt the specific wrong part (i.e., corresponding to the wrong standing posture type) in the case of an incorrect standing posture.

[0049] In step S133, the second training data set DA2 is input into the posture detection model POM, so that the posture detection model POM makes a judgment on each posture image data in the second training data set DA2, and trains the posture detection model POM with reference to the corresponding posture type label. As an example, the posture detection model POM may first make a type judgment on the first posture image data in the second training data set DA2 (for example, the first entry in the second training data set DA2) (for example, judged as a posture type label of a left-hand error), and calculate the difference value (i.e., loss value) between the judgment result and the corresponding type label of the first posture image data. By minimizing the difference value, the parameters of the posture detection model POM are iteratively updated to form an updated posture detection model POM, and the updated posture detection model POM is used to recycle until the loss value meets a certain minimum condition (for example, the loss value is reduced to below a certain threshold, or the number of pre-training is reached). By analogy, the cycle is completed for each posture image data in the second training data set DA2, thereby completing the training process of the posture detection model POM. The posture detection model POM obtained when all cycles of the second training data set DA2 are completed is the trained posture detection model POM'.

[0050] Next, return Figure 1 When the original image data OIM is determined to be human body data in step S3, the process may proceed to step S4. Otherwise, when the original image data OIM is determined to be non-human body data in step S3, the process may proceed to step S5.

[0051] In step S4, the original image data OIM is input into the trained posture detection model POM' to determine whether the human posture is correct as the posture identification result POR. As mentioned above, when it is determined as human body data in step S3, it can indicate that the detected person is in the correct standing position. Therefore, the original image data OIM obtained at this time can continue to be used to detect whether the standing posture is correct. After the posture detection model POM has been trained, it has the ability to accurately identify the image data collected by the image acquisition device for human body security inspection, so the original image data OIM can be input into the trained posture detection model POM' to determine whether the standing posture is correct and possibly prompt the wrong part.

[0052] In one embodiment, in the process of determining whether the human body posture is correct, the trained posture detection model POM' can calculate the confidence for each of the multiple wrong posture types according to the original image data OIM, so as to output the posture discrimination result POR according to the calculated confidence. As an example, multiple wrong posture types can be preset. For example, multiple wrong posture types can be used to train the posture detection model POM. Each wrong posture type can each include at least one representative feature. After receiving the original image data OIM, the trained posture detection model POM' can perform feature extraction on the original image data OIM, and calculate the feature percentage of each wrong posture type as the confidence according to the respective representative features of the multiple wrong posture types. For example, the first wrong posture type includes two representative features. If the features extracted by the trained posture detection model POM' include these two representative features, then 100% or 1 can be calculated. The second wrong posture type includes four representative features. If the features extracted by the trained posture detection model POM' include only one of the four representative features, then 25% or 0.25 can be calculated. Therefore, it can be considered that the confidence of the first wrong stance type is 1, and the confidence of the second wrong stance type is 0.25. However, the above-mentioned method of calculating the confidence is only an example, and the embodiments of the present invention are not limited thereto, and any suitable calculation method can be used.

[0053] Further, the trained posture detection model POM' can compare the calculated confidences of each incorrect posture type (i.e., all the calculated confidences) with the second threshold value respectively. When all confidences are lower than the second threshold value, the human body posture is judged to be correct. When not all confidences are lower than the second threshold value, the human body posture is judged to be wrong, and the incorrect posture type with the highest confidence is used as the posture discrimination result POR. In the example described above, if the trained posture detection model POM' sets the second threshold value to 0.8, the confidence of the first incorrect posture type (which is 1) is higher than the second threshold value, while the confidence of the second incorrect posture type (which is 0.25) is lower than the second threshold value. Therefore, not all confidences are lower than the second threshold value, and it can be determined that the posture of the detected person is wrong. Further, since the confidence of the first incorrect posture type is the highest, the first incorrect posture type can be output as the posture discrimination result POR. For example, if the first incorrect standing posture type is a left-hand error, the external inspection personnel can be prompted that the inspected person has an incorrect left-hand error, that is, the left-hand posture needs to be corrected to be in a correct standing posture. In this way, the inspection personnel can quickly locate the incorrect standing posture of the inspected person and make timely corrections, thereby speeding up the security inspection.

[0054] In step S5, since the original image data OIM is not human body data, this indicates that the detected person may be standing in the wrong position, or the image acquisition device may fail to acquire data, so that the original image data OIM cannot be further used to determine whether the standing posture is correct. Therefore, in step S5, the abnormal image discrimination result IMR that is not human body data can be first output to the outside (for example, the detection personnel) to prompt the detection personnel to adjust the detected person. For example, the external prompt can be given in a visual display or voice prompt. Then, the image acquisition device can be used to re-acquire new original image data OIM, and the new original image data OIM can be re-input into the trained abnormal image detection model IDM' to determine whether it is human body data. For example, the detection personnel can use the image acquisition device to re-acquire new original image data OIM for the detected person. According to the human body security inspection system of the present invention, step S3 can be re-performed on the re-acquired original image data OIM to determine whether it is human body data. Since the execution process of step S3 has been described in detail above, it will not be repeated here. In addition, when the new original image data OIM is determined to be human body data at step S3, step S4 described above can be further performed.

[0055] In step S6, the standing posture identification result POR obtained in step S4 can be output to the outside (for example, the detection personnel). For example, it can be prompted to the outside in a visual display mode or a voice prompt mode, so that the detection personnel can know whether the standing posture of the detected person meets the standard according to the standing posture identification result POR. In one embodiment, in the visual display mode, the human body parts with wrong standing postures can be prompted in the form of puppets without displaying the human body image. This can avoid the risk of human body image data leakage and increase the confidentiality of the data.

[0056] In one embodiment, the abnormal image detection model IDM and the posture detection model POM may be Resnet classification network models. The Resnet classification network model is usually composed of a plurality of residual network structures stacked together. Figure 4 An example of a residual network structure is shown. Figure 4 As shown in the figure, the residual network structure is formed by adding the features obtained from two parts. The first part is the direct input feature x, and the second part is the residual F(x) obtained by performing two to three convolutions on the feature. Finally, the features of the two parts are added together H(x)=F(x)+x.

[0057] According to an embodiment of the present invention, the abnormal image detection model IDM and / or the posture detection model POM is improved on the basis of the Resnet classification network model and is combined with two residual network structures. Figure 5A and Figure 5B Schematic diagrams of two residual network structures combined according to embodiments of the present invention are respectively shown. Figure 5A The residual structure of is composed of two 3x3 convolutional layers in the main branch. Figure 5B The residual structure of is composed of two 3x3 convolutional layers and one 1x1 convolutional layer. The 3x3 convolutional layer is used to learn features and extract feature information, and the 1x1 convolutional layer is used to avoid the gradient disappearance caused by the depth of the network. The residual network can effectively reduce the number of neural network parameters and the complexity of calculation, thereby effectively solving the degradation problem of deep networks.

[0058] Figure 6 A schematic diagram showing the network structure of the Resnet classification network model according to an embodiment of the present invention is shown. Figure 6 As shown in Figure 1, the Resnet classification network model can be composed of an input layer (e.g., an input millimeter wave image), a convolutional layer, four residual layers, a maximum pooling layer, an average pooling layer, a fully connected layer, and an output layer. For example, the first convolutional layer can be connected first, then through a maximum pooling layer, and then the residual structure ( Figure 5A ), then connect the batch normalization layer and the Relu activation function, and then connect 3 residual structures ( Figure 5B ), and finally connect the global average pooling layer and the fully connected layer.

[0059] However, it should be noted that the embodiments of the present invention are not limited to the Resnet classification network model, and other suitable classification network models may be adopted, such as Lenet, Alexnet, VGG, Mobilenet and other classification network models.

[0060] In one embodiment, during the training of the abnormal image detection model IDM and the posture detection model POM, for example, after a training phase, a testing process may be performed on the abnormal image detection model IDM and the posture detection model POM that have passed the training phase. For example, a test data set may be input into the abnormal image detection model IDM and the posture detection model POM that have passed the training phase to test the accuracy of these models. Then, the model may be adjusted and optimized based on the test results to obtain an optimal posture recognition model. In some cases, the test data set may be different from the first training data set DA1 and the second training data set DA2 described above, but may be a separately collected image data set with more dimensions.

[0061] Figure 7 A schematic diagram of a human body security inspection system 100 according to an embodiment of the present invention is shown.

[0062] like Figure 7 As shown, the human body security inspection system 100 may include an image acquisition device 10. The image acquisition device 10 may be configured to acquire original image data OIM within a scanning area.

[0063] The human body security inspection system 100 may further include a computing device 20. For example, the abnormal image detection model IDM and the standing posture detection model POM may be stored in the computing device 20. The computing device 20 may be configured to train the abnormal image detection model IDM to obtain a trained abnormal image detection model IDM'. Then, during the human body security inspection process, the computing device 20 may enable the trained abnormal image detection model IDM' to perform calculations to obtain an abnormal image discrimination result IMR. On the other hand, the computing device 20 may also be configured to train the standing posture detection model POM to obtain a trained standing posture detection model POM'. Then, during the human body security inspection process, the computing device 20 may enable the trained standing posture detection model POM' to perform calculations to obtain a standing posture discrimination result POR.

[0064] The computing device 20 may include a data processing unit 210. The data processing unit 210 may be configured to perform the above-mentioned image preprocessing on the plurality of standing posture image data to form a second training data set DA2 for training the standing posture detection model POM.

[0065] The human body security inspection system 100 may further include a visual display device and / or a voice prompt device 30. The visual display device may be configured to output the abnormal image discrimination result IDM and / or the standing posture discrimination result POR to the outside in a visual display manner. The voice prompt device may be configured to output the abnormal image discrimination result IDM and / or the standing posture discrimination result POR to the outside in a voice prompt manner.

[0066] Advantageously, the human body security inspection method and system according to the embodiments of the present invention can determine whether the posture of the human body is appropriate during security inspection based on the algorithm model of deep learning. The human body security inspection method and system directly processes the original image by integrating image recognition, image processing, and deep learning technology. It is simple and convenient, and has the characteristics of high recognition accuracy, strong generalization performance, and strong stability. It makes up for the defects of large workload and poor applicability of feature extraction in existing image recognition technology. In addition, the human body security inspection method and system according to the embodiments of the present invention can use the trained model for image recognition and judgment processing. In the recognition process, there is no need to use other sensor data, only the directly collected image information (for example, millimeter wave image information) is required, so it can efficiently and accurately identify abnormal images and wrong standing postures, improve security inspection efficiency and user experience.

[0067] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and this document does not limit this.

[0068] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A human body security inspection method, the method comprising: Constructing an abnormal image detection model, and training the abnormal image detection model based on the first training data set to obtain a trained abnormal image detection model; Acquire original image data within the scanning area; Inputting the original image data into a trained abnormal image detection model to determine whether the original image data is human body data as an abnormal image discrimination result; and The abnormal image identification result is outputted to the outside.

2. The method according to claim 1, wherein training the abnormal image detection model based on the first training data set comprises: Collecting a plurality of human body data and a plurality of non-human body data, and adding an attribute category label to each of the collected data, so as to form the first training data set; The first training data set is input into the abnormal image detection model, so that the abnormal image detection model judges each data in the first training data set and performs training with reference to the corresponding attribute category label. 3 . The method according to claim 2 , wherein the plurality of human body data comprises human body data with correct standing posture and human body data with incorrect standing posture, and the plurality of non-human body data comprises a plurality of different types of object data.

4. The method according to claim 2, wherein determining whether the original image data is human body data comprises: Calculating the overall confidence of the original image data, and comparing the overall confidence with a first threshold; as well as When the overall confidence level is greater than the first threshold, it is determined that the original image data is human body data.

5. The method according to claim 1, further comprising: Constructing a standing posture detection model, and training the standing posture detection model based on a second training data set to obtain a trained standing posture detection model; When it is determined that the original image data is human body data, the original image data is input into the trained posture detection model to determine whether the human body posture is correct as a posture determination result; as well as The standing posture identification result is output to the outside.

6. The method according to claim 5, wherein training the posture detection model based on the second training data set comprises: Collecting a plurality of standing posture image data, wherein the plurality of standing posture image data includes standing posture image data of a plurality of wrong standing posture types, and each wrong standing posture type is assigned a standing posture type label; Performing image preprocessing on the plurality of standing posture image data to form the second training data set; as well as The second training data set is input into the standing posture detection model, so that the standing posture detection model judges each standing posture image data in the second training data set, and trains the standing posture detection model with reference to the corresponding standing posture type label.

7. The method according to claim 6, wherein performing image preprocessing on the plurality of standing posture image data comprises at least one of the following: Performing screening processing on the collected multiple standing posture image data to delete invalid standing posture image data; Performing a uniformization process on the collected multiple standing posture image data to unify the number of standing posture image data of each wrong standing posture type; Performing expansion processing on the collected multiple standing posture image data; Performing normalization processing on the collected multiple standing posture image data; and A randomization process is performed on the collected plurality of standing posture image data.

8. The method according to claim 6, wherein determining whether the human body posture is correct comprises: Calculating confidence respectively for each of the multiple incorrect standing posture types according to the original image data; When all the calculated confidence levels are lower than the second threshold, it is determined that the human body's standing posture is correct; When all the calculated confidences are not lower than the second threshold, it is determined that the human body's standing posture is wrong, and the wrong standing posture type with the highest confidence is used as the standing posture determination result.

9. The method according to claim 5, further comprising: When it is determined that the original image data is not the human body data, new original image data is recaptured, and the new original image data is re-input into the trained abnormal image detection model to determine whether it is human body data.

10. The method according to any one of claims 5 to 9, further comprising: The abnormal image identification result and / or the standing posture identification result are output to the outside in a visual display mode or a voice prompt mode to provide an error prompt during the human body security inspection.

11. According to the method according to any one of claims 5-9, the abnormal image detection model and the standing posture detection model are Resnet classification network models.

12. A human body security inspection system, the system executing the human body security inspection method according to any one of claims 1 to 11, the system comprising: An image acquisition device, wherein the image acquisition device is configured to acquire raw image data within a scanning area; as well as A computing device is configured to train an abnormal image detection model to obtain a trained abnormal image detection model, and to enable the trained abnormal image detection model to perform calculations to obtain abnormal image discrimination results.

13. The system according to claim 12, wherein: The computing device is also configured to train the stance detection model to obtain a trained stance detection model, and to enable the trained stance detection model to perform calculations to obtain a stance discrimination result.

14. The system of claim 13, further comprising: A visual display device, wherein the visual display device is configured to output the abnormal image discrimination result and / or the standing posture discrimination result to the outside in a visual display manner; and / or A voice prompt device is configured to output the abnormal image discrimination result and / or the standing posture discrimination result to the outside in the form of voice prompt.

15. The system of claim 13, wherein: The image acquisition device is further configured to acquire a plurality of standing posture image data; and The computing device also includes a data processing unit, which is configured to perform image preprocessing on the multiple standing posture image data to form a second training data set for training the standing posture detection model.

Images