Weak feature background visual risk control method based on image deep analysis

Abstract

The invention relates to a weak feature background visual risk control method based on image deep analysis. The method comprises the following steps: acquiring a to-be-detected image; inputting the to-be-detected image into a multi-scale convolutional neural network so as to extract multi-scale features of a background region from the to-be-detected image through the multi-scale convolutional neural network; calculating the feature change difference of the background region based on the multi-scale features through a multi-scale convolutional neural network; and identifying a weak feature background in the background region through a multi-scale convolutional neural network according to the feature change difference. According to the scheme provided by the invention, weak feature backgrounds can be automatically and accurately identified, manual intervention is reduced, the robustness and accuracy of similar background detection are improved, and the method is suitable for various complex scenes and has relatively strong generalization ability.

Description

Technical Field

[0001] The present application relates to the field of remote risk control technology, and in particular to a weak feature background visual risk control method based on image depth analysis. Background Art

[0002] Similar background detection refers to finding and marking image areas that are similar to the background but have specific targets in a group of images. Similar background detection can effectively reduce the risk of collective fraud.

[0003] In similar background detection, weak feature backgrounds (such as white walls and single-color backgrounds) are often easily mistaken for having a high similarity with the target background due to the lack of obvious texture or feature information. In this case, the detection system may misclassify or over-aggregate the weak feature background, resulting in deviations in similar background detection results and even affecting the execution of subsequent tasks.

[0004] In the related art, it is necessary to manually predefine the weak feature background type, which not only increases the complexity of the operation but also makes it difficult to adapt to a variety of complex scenes. In addition, the related art relies on matching a single feature (such as texture), which is not sensitive enough to weak feature backgrounds such as white walls and is easily disturbed by complex backgrounds, which in turn easily leads to over-aggregation of weak feature backgrounds and is insufficient to cover all weak feature background scenes, resulting in low recognition accuracy. Summary of the invention

[0005] In order to solve or partially solve the problems existing in the related technology, the present application provides a weak feature background visual risk control method based on image deep analysis, which can automatically and accurately identify weak feature backgrounds, reduce manual intervention, improve the robustness and accuracy of similar background detection, is suitable for a variety of complex scenes, and has strong generalization ability.

[0006] The first aspect of the present application provides a weak feature background visual risk control method based on image depth analysis, the method comprising: Acquire the image to be detected; Inputting the image to be detected into a multi-scale convolutional neural network to extract multi-scale features of a background area from the image to be detected through the multi-scale convolutional neural network; Calculating the feature change difference of the background area based on the multi-scale features by the multi-scale convolutional neural network; The multi-scale convolutional neural network is used to identify the weak feature background in the background area according to the feature change difference.

[0007] In one embodiment, the method further comprises: Eliminating the weak feature background from the background area to obtain a non-weak feature area; Perform similar background detection on the non-weak feature area.

[0008] In one embodiment, the multi-scale convolutional neural network includes a first convolution kernel and a second convolution kernel of different sizes, and the multi-scale features include large-scale features and small-scale features; and extracting the multi-scale features of the background area from the image to be detected by the multi-scale convolutional neural network includes: Inputting the image to be detected into the first convolution kernel and the second convolution kernel respectively through the multi-scale convolution neural network; Extracting a plurality of the large-scale features of the background area from the image to be detected by using the first convolution kernel; and A plurality of the small-scale features of the background area are extracted from the image to be detected by using the second convolution kernel.

[0009] In one embodiment, the extracting the plurality of large-scale features of the background area from the image to be detected by using the first convolution kernel includes: Dividing the background area in the image to be detected into a plurality of first blocks by using the first convolution kernel; Extracting feature information of each of the first blocks by using the first convolution kernel to obtain each of the large-scale features of the background area; The extracting a plurality of the small-scale features of the background area from the image to be detected by using the second convolution kernel includes: Dividing the background area in the image to be detected into a plurality of second blocks by using the second convolution kernel; Extracting feature information of each of the second blocks by using the second convolution kernel to obtain each of the small-scale features of the background area; The size of the first block is larger than that of the second block, and the feature information of the first block and the feature information of the second block both include texture, color and shape.

[0010] In one embodiment, the calculating the feature change difference of the background area based on the multi-scale features by the multi-scale convolutional neural network includes: Calculating the feature change difference between two adjacent first blocks based on feature information of two adjacent first blocks by using the first convolution kernel; and The feature change difference between two adjacent second blocks is calculated based on the feature information of two adjacent second blocks by using the second convolution kernel.

[0011] In one embodiment, identifying the weak feature background in the background area according to the feature change difference by the multi-scale convolutional neural network includes: Comparing the feature change difference between two adjacent first blocks with a first change threshold through the first convolution kernel; and Compare the feature change difference between two adjacent second blocks with a second change threshold using the second convolution kernel; If the feature change difference between two adjacent first blocks is less than or equal to the first change threshold, a first recognition result output by the first convolution kernel is obtained; the first recognition result is used to characterize that the two adjacent first blocks constitute a first weak feature area; and / or, If the feature change difference between two adjacent second blocks is less than or equal to the second change threshold, a second recognition result output by the second convolution kernel is obtained; the second recognition result is used to indicate that the two adjacent second blocks constitute a second weak feature area; A sub-block commonly included in the first weak feature region and the second weak feature region is identified as a weak feature background in the background region.

[0012] In one embodiment, the step of acquiring the image to be detected includes: During the remote interaction with the lender, collecting visual data of the scene in which the lender is located; wherein the visual data includes at least one of image data and video data; The visual data is preprocessed to obtain an image to be detected; wherein the preprocessing operation includes at least one of image denoising and image normalization.

[0013] The second aspect of the present application provides a weak feature background visual wind control system based on image depth analysis, the system comprising: An image acquisition module, used for acquiring an image to be detected; A feature extraction module, used for inputting the image to be detected into a multi-scale convolutional neural network, so as to extract multi-scale features of a background area from the image to be detected through the multi-scale convolutional neural network; A feature change difference calculation module, used to calculate the feature change difference of the background area based on the multi-scale feature through the multi-scale convolutional neural network; A weak feature background recognition module is used to identify the weak feature background in the background area according to the feature change difference through the multi-scale convolutional neural network.

[0014] A third aspect of the present application provides an electronic device, including: Processor; and The memory stores executable codes thereon, and when the executable codes are executed by the processor, the processor is caused to execute the method as described above.

[0015] A fourth aspect of the present application provides a computer-readable storage medium having executable code stored thereon. When the executable code is executed by a processor of an electronic device, the processor is caused to execute the method as described above.

[0016] The technical solution provided by this application may have the following beneficial effects: The solution provided by the present application is to obtain an image to be detected; input the image to be detected into a multi-scale convolutional neural network to extract multi-scale features of the background area from the image to be detected through the multi-scale convolutional neural network; calculate the feature change difference of the background area based on the multi-scale features through the multi-scale convolutional neural network; and identify the weak feature background in the background area according to the feature change difference through the multi-scale convolutional neural network. The present application uses a multi-scale convolutional neural network to automatically identify the weak feature background, thereby reducing manual intervention and improving recognition efficiency. The multi-scale convolutional neural network can extract multi-scale features of the background area, and the multi-scale features can accurately reflect the feature change differences of the background area, and then effectively identify the weak feature background based on the feature change differences, thereby improving the robustness and accuracy of similar background detection. It is also applicable to a variety of complex scenes and has strong generalization capabilities.

[0017] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above and other objects, features and advantages of the present application will become more apparent by describing in more detail the exemplary embodiments of the present application in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments of the present application.

[0019] Figure 1 It is a flow chart of a visual wind control method for weak feature background based on image depth analysis shown in an embodiment of the present application; Figure 2 It is another flow chart of the weak feature background visual wind control method based on image depth analysis shown in the embodiment of the present application; Figure 3 It is a structural schematic diagram of a weak feature background visual wind control system based on image depth analysis shown in an embodiment of the present application; Figure 4 It is a schematic diagram of the structure of an electronic device shown in an embodiment of the present application. DETAILED DESCRIPTION

[0020] The embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although the embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to make the present application more thorough and complete, and to fully convey the scope of the present application to those skilled in the art.

[0021] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The singular forms of "a", "said" and "the" used in this application and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0022] It should be understood that although the terms "first", "second", "third", etc. may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of this application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of this application, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0023] Similar background detection can effectively reduce the risk of collusion. Specifically, during the loan application process, the borrower needs to pass facial recognition authentication, in which the background area can reflect the scene the borrower is in. By matching the features (such as texture) of the entire background area with the features (such as texture) of the target background (pre-collected background with collusion risk), if the two match, it is determined that the background area is similar to the target background, and then the scene where the borrower is in is assessed to have the risk of collusion.

[0024] However, if there is a weak feature background (such as a white wall or a single color background) in the background area, it is easy to be mistaken for having a high similarity with the target background due to the lack of obvious texture or feature information. For example, suppose that there are multiple borrowers applying for loans in different places within a short period of time, but their backgrounds are all white walls. This can easily cause the detection system to determine that these borrowers are in the same place, and then mistakenly aggregate these irrelevant borrowers together, assessing that there is a risk of collusion in the scene they are in. It can be seen that weak feature backgrounds can easily cause deviations in similar background detection results and even affect the execution of subsequent tasks. Therefore, it is necessary to identify weak feature backgrounds to reduce their interference with similar background detection.

[0025] In the related art, weak feature backgrounds need to be identified manually, which requires the type of weak feature backgrounds to be predefined. However, this not only increases the complexity of the operation but also makes it difficult to adapt to a variety of complex scenes. In addition, the related art relies on matching a single feature (such as texture), which is not sensitive enough to weak feature backgrounds such as white walls and is easily disturbed by complex backgrounds, which in turn easily leads to over-aggregation of weak feature backgrounds and is insufficient to cover all weak feature background scenes, resulting in low recognition accuracy.

[0026] In response to the above problems, the embodiments of the present application provide a weak feature background visual risk control method based on image depth analysis, which uses a multi-scale convolutional neural network to automatically identify weak feature backgrounds, thereby reducing manual intervention and improving recognition efficiency. The multi-scale convolutional neural network can extract multi-scale features of the background area, and the multi-scale features can accurately reflect the feature change differences of the background area, and then effectively identify weak feature backgrounds based on the feature change differences, thereby improving the robustness and accuracy of similar background detection. It is also suitable for a variety of complex scenes and has strong generalization capabilities.

[0027] The technical solution of the embodiments of the present application is described in detail below with reference to the accompanying drawings.

[0028] Figure 1 It is a flow chart of a weak feature background visual risk control method based on image depth analysis shown in an embodiment of the present application.

[0029] See also Figure 1 The weak feature background visual risk control method based on image depth analysis of the present application includes: S110, obtaining an image to be detected.

[0030] In the embodiment of the present application, it can be applied to a weak feature background visual risk control system based on image depth analysis (hereinafter referred to as the "visual risk control system"). When a user needs a video face-to-face signature or a loan application, the user acts as a lender, and the visual risk control system can obtain visual data of the scene in which the lender is located. The visual data may include at least one of static image data and dynamic video data. Since similar background detection is essentially an analysis of background areas other than people (such as lenders), the visual risk control system can perform portrait segmentation on the visual data to obtain an image to be detected, and the image to be detected may only contain background areas other than people (such as lenders).

[0031] S120, inputting the image to be detected into a multi-scale convolutional neural network to extract multi-scale features of the background area from the image to be detected through the multi-scale convolutional neural network.

[0032] The embodiment of the present application can pre-train a multi-scale convolutional neural network (MSCNN), which is a deep learning model for processing features of different scales. The embodiment of the present application can deploy the trained multi-scale convolutional neural network into the visual risk control system for online use by the visual risk control system.

[0033] In practical applications, the visual risk control system can call a pre-trained multi-scale convolutional neural network, and then input the image to be detected into the multi-scale convolutional neural network, so as to extract multi-scale features of the background area from the image to be detected through the multi-scale convolutional neural network.

[0034] It should be noted that, compared to the traditional convolutional neural network that only performs convolution and pooling operations at one scale, the embodiment of the present application utilizes a multi-scale convolutional neural network (MSCNN) that can operate at different scales, specifically by constructing multiple parallel convolution paths, each path using convolution kernels of different sizes to capture feature information at different scales. For example, in a multi-scale convolutional neural network, a smaller convolution kernel can capture local microscopic features of the background area from the image to be detected, and a larger convolution kernel can capture global macroscopic features of the background area from the image to be detected. Therefore, the multi-scale features extracted by the embodiment of the present application can cover the details and overall information of the background area.

[0035] S130, calculating feature change differences of the background area based on the multi-scale features through a multi-scale convolutional neural network.

[0036] After extracting the multi-scale features of the background area, the multi-scale convolutional neural network can calculate the feature change difference of the background area based on the multi-scale features. The feature change difference is used to characterize the degree of difference in feature information between two adjacent blocks in the same background area.

[0037] S140, identifies weak feature background in the background area according to feature change differences through a multi-scale convolutional neural network.

[0038] In one example, if the feature change difference is small or even close to zero, it means that the feature information of the two adjacent blocks in the background area has not changed significantly or has not changed at all, then the multi-scale convolutional neural network can identify the two adjacent blocks as weak feature backgrounds in the background area. The weak feature backgrounds may include but are not limited to: white walls, single color backgrounds, and backgrounds with common decoration styles.

[0039] In another example, if the feature change difference is large, it means that the feature information of two adjacent blocks in the background area has changed significantly. Then the multi-scale convolutional neural network can identify the two adjacent blocks as non-weak feature areas in the background area, so that the visual risk control system can subsequently focus on analyzing the non-weak feature areas, thereby greatly improving the recognition ability in complex backgrounds.

[0040] As can be seen from this example, the solution provided by this application is to obtain an image to be detected; input the image to be detected into a multi-scale convolutional neural network to extract multi-scale features of the background area from the image to be detected through the multi-scale convolutional neural network; calculate the feature change difference of the background area based on the multi-scale features through the multi-scale convolutional neural network; and identify the weak feature background in the background area according to the feature change difference through the multi-scale convolutional neural network. This application uses a multi-scale convolutional neural network to automatically identify the weak feature background, thereby reducing manual intervention and improving recognition efficiency. The multi-scale convolutional neural network can extract multi-scale features of the background area, and the multi-scale features can accurately reflect the feature change differences of the background area, and then effectively identify the weak feature background based on the feature change differences, thereby improving the robustness and accuracy of similar background detection. It is also suitable for a variety of complex scenes and has strong generalization capabilities.

[0041] Figure 2 It is another flow chart of the weak feature background visual risk control method based on image depth analysis shown in this application.

[0042] See also Figure 2 The weak feature background visual risk control method based on image depth analysis of the present application includes: S210, obtaining an image to be detected.

[0043] For this step, please refer to the description in S110 and will not be repeated here.

[0044] In one implementation, acquiring the image to be detected may include: During the remote interaction with the lender, visual data of the scene in which the lender is located is collected; wherein the visual data includes at least one of image data and video data; a preprocessing operation is performed on the visual data to obtain an image to be detected; wherein the preprocessing operation includes at least one of image denoising and image normalization.

[0045] When a user needs a video interview or loan application, the user acts as a lender. The lender can initiate a remote interaction request to the visual risk control system through a terminal device (such as a mobile phone, computer, etc.). The visual risk control system responds to the remote interaction request and interacts remotely with the lender. The visual risk control system can perform visual risk control on the lender's scene in the current remote interaction scenario.

[0046] In one example, after receiving a remote interaction request, the visual risk control system can send a collection instruction to the terminal device currently used by the lender. The collection instruction is used to instruct the camera of the terminal device to collect visual data of the scene in which the lender is located in real time. Therefore, during the remote interaction process with the lender, the visual risk control system can obtain visual data of the scene in which the lender is located in real time.

[0047] In another example, the remote interaction request may carry visual data of the scene uploaded by the lender, so that during the remote interaction process with the lender, the visual risk control system can obtain the visual data of the scene where the lender is located from the remote interaction request.

[0048] The visual data may include at least one of static image data and dynamic video data, and both the image data and the video data may be in RGB format.

[0049] In order to enhance the deep representation of weak features, the visual risk control system can perform preprocessing operations such as image denoising and image normalization on the visual data to obtain the image to be detected. Among them, image denoising can remove noise interference. For example, in a scene with large noise, the visual risk control system can use Gaussian filtering to remove noise interference. Image normalization is through a series of transformations (i.e., using the invariant moments of visual data to find a set of parameters that can eliminate the influence of other transformation functions on the transformation of visual data), so that the visual data is converted into a corresponding unique standard form of the image to be detected. The image to be detected can have invariant characteristics to affine transformations such as translation, rotation, and scaling.

[0050] In addition, since similar background detection is essentially an analysis of background areas other than people (such as lenders), the visual risk control system can also perform portrait segmentation on visual data to obtain an image to be detected, which may only contain background areas other than people (such as lenders).

[0051] S220, inputting the image to be detected into a multi-scale convolutional neural network to extract multi-scale features of the background area from the image to be detected through the multi-scale convolutional neural network.

[0052] For this step, please refer to the description in S120 and will not be repeated here.

[0053] In one embodiment, the multi-scale convolutional neural network includes a first convolution kernel and a second convolution kernel of different sizes, and the multi-scale features include large-scale features and small-scale features; extracting the multi-scale features of the background area from the image to be detected by the multi-scale convolutional neural network may include: The image to be detected is input into the first convolution kernel and the second convolution kernel respectively through a multi-scale convolutional neural network; a number of large-scale features of the background area are extracted from the image to be detected through the first convolution kernel; and a number of small-scale features of the background area are extracted from the image to be detected through the second convolution kernel.

[0054] The embodiment of the present application can pre-construct multiple parallel convolution paths in a multi-scale convolutional neural network (MSCNN), and each path uses convolution kernels of different sizes, so that the multi-scale convolutional neural network can include convolution kernels of different sizes, so as to use convolution kernels of different sizes to extract multi-scale features of the background area. Specifically, the convolution kernels of different sizes include at least a relatively large first convolution kernel and a relatively small second convolution kernel, and the multi-scale features of the background area include at least large-scale features and small-scale features. The multi-scale convolutional neural network inputs the image to be detected into the first convolution kernel and the second convolution kernel through two parallel convolution paths respectively. Since the size of the first convolution kernel is larger than the size of the second convolution kernel, the first convolution kernel is responsible for extracting several large-scale features of the background area from the image to be detected, and the second convolution kernel is responsible for extracting several small-scale features of the background area from the image to be detected.

[0055] Among them, large-scale features are used to characterize the global macroscopic features of the background area, and small-scale features are used to characterize the local microscopic features of the background area. As an example, assuming that the color of the background area changes slowly, if only small-scale features are extracted, the color basically does not change from a microscopic point of view, but from a macroscopic point of view, a slow change in color can be found. As another example, assuming that the color of a small block in the background area changes drastically, if only large-scale features are extracted, the detailed information of the small block is easily lost from a macroscopic point of view, but from a microscopic point of view, it can be found that the color of the small block changes drastically. Therefore, the embodiments of the present application can cover the details and overall information of the background area by combining large-scale features with small-scale features.

[0056] In one embodiment, extracting several large-scale features of the background area from the image to be detected by the first convolution kernel may include: The background area in the image to be detected is divided into a plurality of first blocks by using a first convolution kernel; the feature information of each first block is extracted by using the first convolution kernel to obtain various large-scale features of the background area; wherein the feature information of the first block includes texture, color and shape.

[0057] Since similar background detection is essentially an analysis of background areas other than people (such as lenders), and the image to be detected only contains background areas other than people (such as lenders), the first convolution kernel can divide the background area in the image to be detected into a number of first blocks, and then extract the feature information of each first block, and the feature information of each first block can be used as a large-scale feature of the background area. The feature information of each first block can include texture, color and shape.

[0058] In one embodiment, extracting several small-scale features of the background area from the image to be detected by the second convolution kernel may include: The background area in the image to be detected is divided into a plurality of second blocks by a second convolution kernel; the feature information of each second block is extracted by the second convolution kernel to obtain various small-scale features of the background area; wherein the feature information of the second block includes texture, color and shape.

[0059] Since similar background detection is essentially an analysis of background areas other than people (such as lenders), and the image to be detected only contains background areas other than people (such as lenders), the second convolution kernel can divide the background area in the image to be detected into a number of second blocks, and then extract the feature information of each second block, and the feature information of each second block can be used as a small-scale feature of the background area. The feature information of each second block can include texture, color and shape.

[0060] It should be noted that, since the size of the first convolution kernel is larger than the size of the second convolution kernel, the size of the first block is larger than the size of the second block.

[0061] S230, calculating feature change differences of the background area based on the multi-scale features through a multi-scale convolutional neural network.

[0062] For this step, please refer to the description in S130 and will not be repeated here.

[0063] In one embodiment, calculating the feature change difference of the background area based on the multi-scale features by a multi-scale convolutional neural network may include: The feature change difference between two adjacent first blocks is calculated based on the feature information of the two adjacent first blocks by using the first convolution kernel; and the feature change difference between two adjacent second blocks is calculated based on the feature information of the two adjacent second blocks by using the second convolution kernel.

[0064] The difference in feature changes can be expressed by the change in feature mean or feature variance.

[0065] In one example, after the background area is divided into multiple first blocks, the first convolution kernel can calculate the texture mean, color mean and shape mean of all pixels in the same first block, and then add the texture mean, color mean and shape mean to obtain the feature mean of the first block. In this way, the feature mean of each first block is obtained, and then based on the feature means of two adjacent first blocks, the feature mean change of two adjacent first blocks is calculated, and the feature mean change can be used as the feature change difference of the two adjacent first blocks.

[0066] In another example, after the background area is divided into multiple first blocks, the first convolution kernel can calculate the texture variance, color variance, and shape variance of all pixels in the same first block, and then add the texture variance, color variance, and shape variance to obtain the feature variance of the first block. In this way, the feature variance of each first block is obtained, and then based on the feature variance of two adjacent first blocks, the feature variance change of two adjacent first blocks is calculated, and the feature variance change can be used as the feature change difference of the two adjacent first blocks.

[0067] Similarly, after the background area is divided into multiple second blocks, the second convolution kernel can calculate the texture mean, color mean and shape mean of all pixels in the same second block, and then add the texture mean, color mean and shape mean to obtain the feature mean of the second block. In this way, the feature mean of each second block is obtained, and then based on the feature mean of two adjacent second blocks, the feature mean change of two adjacent second blocks is calculated, and the feature mean change can be used as the feature change difference of the two adjacent second blocks.

[0068] Similarly, after the background area is divided into multiple second blocks, the second convolution kernel can calculate the texture variance, color variance and shape variance of all pixels in the same second block, and then add the texture variance, color variance and shape variance to obtain the feature variance of the second block. In this way, the feature variance of each second block is obtained, and then based on the feature variance of two adjacent second blocks, the feature variance change of two adjacent second blocks is calculated, and the feature variance change can be used as the feature change difference of the two adjacent second blocks.

[0069] It should be noted that the first convolution kernel and the second convolution kernel can be processed in parallel, thereby improving computational efficiency.

[0070] S240, identifies weak feature background in the background area based on feature change differences through a multi-scale convolutional neural network.

[0071] For this step, please refer to the description in S140 and will not be repeated here.

[0072] In one embodiment, identifying weak feature background in the background area according to feature change differences through a multi-scale convolutional neural network may include: The feature change difference between two adjacent first blocks is compared with a first change threshold through a first convolution kernel; and the feature change difference between two adjacent second blocks is compared with a second change threshold through a second convolution kernel; if the feature change difference between two adjacent first blocks is less than or equal to the first change threshold, a first recognition result output by the first convolution kernel is obtained; the first recognition result is used to characterize that the two adjacent first blocks constitute a first weak feature area; and / or, if the feature change difference between two adjacent second blocks is less than or equal to the second change threshold, a second recognition result output by the second convolution kernel is obtained; the second recognition result is used to characterize that the two adjacent second blocks constitute a second weak feature area; the sub-block commonly contained in the first weak feature area and the second weak feature area is identified as a weak feature background in the background area.

[0073] The embodiment of the present application can identify potential weak feature background in the background area based on the change threshold.

[0074] In one example, the first convolution kernel can compare the feature change difference between two adjacent first blocks with a preset first change threshold. If the feature change difference between two adjacent first blocks is ≤ the first change threshold, the first convolution kernel outputs a first recognition result for the two adjacent first blocks, and the first recognition result is used to characterize that the two adjacent first blocks constitute a first weak feature area. Alternatively, if the feature change difference between two adjacent first blocks is > the first change threshold, the first convolution kernel outputs a third recognition result for the two adjacent first blocks, and the third recognition result is used to characterize that the two adjacent first blocks constitute a first non-weak feature area.

[0075] In another example, the second convolution kernel can compare the feature change difference between two adjacent second blocks with a preset second change threshold. If the feature change difference between two adjacent second blocks is ≤ the second change threshold, the second convolution kernel outputs a second recognition result for the two adjacent second blocks, and the second recognition result is used to characterize that the two adjacent second blocks constitute a second weak feature area. If the feature change difference between two adjacent second blocks is > the second change threshold, the second convolution kernel outputs a fourth recognition result for the two adjacent second blocks, and the fourth recognition result is used to characterize that the two adjacent second blocks constitute a second non-weak feature area.

[0076] Afterwards, the embodiment of the present application may identify the sub-block commonly included in the first weak feature region and the second weak feature region as the weak feature background in the background region.

[0077] In one example, if the first recognition result and the second recognition result are obtained, the visual wind control system may first determine the sub-block commonly contained in the first weak feature area targeted by the first recognition result and the second weak feature area targeted by the second recognition result, so as to identify the sub-block as a weak feature background in the background area. Exemplarily, assuming that the first block A and the second block B jointly contain a sub-block C, if the feature change difference between the first block A and its adjacent first block A' is ≤ the first change threshold, and the feature change difference between the second block B and its adjacent second block B' is ≤ the second change threshold, then the sub-block C can be identified as a weak feature background in the background area.

[0078] In addition, the embodiment of the present application can identify the sub-block commonly included in the first non-weak feature area and the second non-weak feature area as a non-weak feature area in the background area; or, the embodiment of the present application can identify the sub-block commonly included in the first non-weak feature area and the second non-weak feature area as a non-weak feature area in the background area; or, the embodiment of the present application can identify the sub-block commonly included in the first non-weak feature area and the second non-weak feature area as a non-weak feature area in the background area.

[0079] It can be seen that compared with the related art that only uses a single feature (such as texture) to identify weak feature background, the embodiment of the present application improves the recognition accuracy by integrating all features (such as texture, color and shape) to identify weak feature background.

[0080] It can be seen that compared with the related art of analyzing the features of the entire background area, the embodiment of the present application divides the entire background area into large-scale features and small-scale features for analysis, which can cover the details and overall information of the background area, thereby further improving the recognition accuracy.

[0081] S250, removing the weak feature background from the background area to obtain a non-weak feature area.

[0082] The visual risk control system can mark the weak feature background as low priority, and then remove the weak feature background from the background area. The removed background area can be used as a non-weak feature area.

[0083] Compared with the related art in which high-precision similar background detection usually relies on expensive hardware equipment, the embodiment of the present application eliminates weak feature background so that the weak feature background does not need to participate in subsequent similar background detection (i.e., step S260). This can reduce the participation of invalid areas (i.e., weak feature background), significantly reduce the hardware requirements and computing burden of the visual wind control system, thereby reducing computing costs, and can also eliminate the interference of weak feature background on similar background detection, thereby improving detection accuracy.

[0084] In addition, the embodiment of the present application can adjust the degree of discarding the weak feature background by adjusting the first change threshold and the second change threshold. As an example, the first change threshold and the second change threshold can be increased so that the background as long as the feature change is not very strong can be discarded, which can reduce the hardware resources required for subsequent similar background detection and reduce the computing cost. As another example, the first change threshold and the second change threshold can be reduced so that only the background with very smooth feature changes can be discarded, which can increase the hardware resources required for subsequent similar background detection and optimize system performance. The embodiment of the present application can achieve a dynamic balance between resources and performance by adjusting parameters based on the difference in hardware resources.

[0085] In addition, the related art lacks an analytical method that effectively combines depth information to identify weak feature backgrounds, resulting in insufficient expressiveness of weak features in three-dimensional scenes. In response to this, the embodiments of the present application can achieve efficient recognition and classification of weak features in complex backgrounds by combining the depth information of the image on the basis of feature extraction technology. Specifically, before eliminating weak feature backgrounds, the visual wind control system can first obtain the depth information of two adjacent first blocks and the depth information of two adjacent second blocks collected by the ambient light sensor or laser ranging equipment, and then further optimize the elimination results based on the comparison of the depth information of the two adjacent first blocks and the comparison of the depth information of the two adjacent second blocks, thereby further improving the recognition accuracy and efficiency of weak feature backgrounds.

[0086] S260, performing similar background detection on the non-weak feature area.

[0087] The visual risk control system can perform similar background detection on the non-weak feature area. Specifically, the embodiment of the present application pre-collects various target backgrounds with a risk of collusion, and then stores these target backgrounds in the base library of the visual risk control system, so that in this step, various target backgrounds can be obtained from the base library, and then the non-weak feature area is matched with the various target backgrounds respectively. If the non-weak feature area matches one of the target backgrounds, the visual risk control system can determine that the scene where the borrower is located has a risk of collusion, and therefore the loan application of the borrower can be rejected; if the non-weak feature area does not match any of the target backgrounds, the visual risk control system can determine that the scene where the borrower is located does not have a risk of collusion, and therefore the loan application of the borrower can be approved.

[0088] As can be seen from this example, the solution provided by the present application is to obtain an image to be detected; input the image to be detected into a multi-scale convolutional neural network to extract multi-scale features of the background area from the image to be detected through the multi-scale convolutional neural network; calculate the feature change difference of the background area based on the multi-scale features through the multi-scale convolutional neural network; identify the weak feature background in the background area according to the feature change difference through the multi-scale convolutional neural network; remove the weak feature background from the background area to obtain a non-weak feature area; perform similar background detection on the non-weak feature area. The present application uses a multi-scale convolutional neural network to automatically identify the weak feature background, thereby reducing manual intervention and improving recognition efficiency, and the multi-scale convolutional neural network can extract multi-scale features of the background area, and the multi-scale features accurately reflect the feature change difference of the background area, and then effectively identify the weak feature background based on the feature change difference, thereby improving the robustness and accuracy of similar background detection, and is also applicable to a variety of complex scenes, with strong generalization ability. Further, the present application can significantly reduce the hardware requirements and computational burden by reducing the participation of invalid areas (i.e., weak feature backgrounds), so that only the non-weak feature areas need to be analyzed in a focused manner, which can greatly improve the recognition ability in complex backgrounds.

[0089] Corresponding to the aforementioned application function implementation method embodiment, the present application also provides a weak feature background visual wind control system based on image depth analysis, an electronic device, a computer-readable storage medium and corresponding embodiments.

[0090] Figure 3 It is a structural schematic diagram of a weak feature background visual wind control system based on image depth analysis shown in an embodiment of the present application.

[0091] See also Figure 3 The present application provides a weak feature background visual wind control system based on image depth analysis, which may include: An image acquisition module 310 is used to acquire an image to be detected; A feature extraction module 320 is used to input the image to be detected into a multi-scale convolutional neural network to extract multi-scale features of the background area from the image to be detected through the multi-scale convolutional neural network; A feature change difference calculation module 330 is used to calculate the feature change difference of the background area based on the multi-scale features through a multi-scale convolutional neural network; The weak feature background recognition module 340 is used to recognize the weak feature background in the background area according to the feature change difference through a multi-scale convolutional neural network.

[0092] In one embodiment, the system may further include: A removal module is used to remove the weak feature background from the background area to obtain a non-weak feature area; The similar background detection module is used to perform similar background detection on non-weak feature areas.

[0093] In one embodiment, the multi-scale convolutional neural network includes a first convolution kernel and a second convolution kernel of different sizes, and the multi-scale features include large-scale features and small-scale features; the feature extraction module 320 may include: An input submodule, used to input the image to be detected into the first convolution kernel and the second convolution kernel respectively through a multi-scale convolutional neural network; A large-scale feature extraction submodule, used to extract a plurality of large-scale features of the background area from the image to be detected by using a first convolution kernel; and The small-scale feature extraction submodule is used to extract a plurality of small-scale features of the background area from the image to be detected through the second convolution kernel.

[0094] In one embodiment, the large-scale feature extraction submodule may include: A first block division unit, used for dividing the background area in the image to be detected into a plurality of first blocks by using a first convolution kernel; A feature information extraction unit for the first block, used to extract feature information of each first block through a first convolution kernel to obtain each large-scale feature of the background area; The small-scale feature extraction submodule may include: A second block division unit, used for dividing the background area in the image to be detected into a plurality of second blocks by using a second convolution kernel; A feature information extraction unit for the second block, used to extract feature information of each second block through a second convolution kernel to obtain each small-scale feature of the background area; The size of the first block is larger than that of the second block, and the feature information of the first block and the feature information of the second block both include texture, color and shape.

[0095] In one embodiment, the feature change difference calculation module 330 may include: A large-scale feature change difference calculation submodule, used to calculate the feature change difference between two adjacent first blocks based on feature information of two adjacent first blocks through a first convolution kernel; and The small-scale feature change difference calculation submodule is used to calculate the feature change difference between two adjacent second blocks based on the feature information of the two adjacent second blocks through the second convolution kernel.

[0096] In one embodiment, the weak feature background recognition module 340 may include: A large-scale feature change difference comparison submodule is used to compare the feature change difference between two adjacent first blocks with a first change threshold through a first convolution kernel, and, A small-scale feature change difference comparison submodule, used for comparing the feature change difference between two adjacent second blocks with a second change threshold through a second convolution kernel; A first recognition result obtaining submodule is used to obtain a first recognition result output by a first convolution kernel if the feature change difference between two adjacent first blocks is less than or equal to a first change threshold; the first recognition result is used to characterize that the two adjacent first blocks constitute a first weak feature area; and, A second recognition result obtaining submodule is used to obtain a second recognition result output by a second convolution kernel if the feature change difference between two adjacent second blocks is less than or equal to a second change threshold; the second recognition result is used to characterize that the two adjacent second blocks constitute a second weak feature area; The weak feature background recognition submodule is used to recognize the sub-block contained in the first weak feature area and the second weak feature area as the weak feature background in the background area.

[0097] In one embodiment, the image acquisition module 310 may include: A visual data collection submodule, used to collect visual data of the scene where the lender is located during the remote interaction with the lender; wherein the visual data includes at least one of image data and video data; The preprocessing operation submodule is used to perform preprocessing operations on visual data to obtain an image to be detected; wherein the preprocessing operation includes at least one of image denoising and image normalization.

[0098] As can be seen from this example, the solution provided by this application is to obtain an image to be detected; input the image to be detected into a multi-scale convolutional neural network to extract multi-scale features of the background area from the image to be detected through the multi-scale convolutional neural network; calculate the feature change difference of the background area based on the multi-scale features through the multi-scale convolutional neural network; and identify the weak feature background in the background area according to the feature change difference through the multi-scale convolutional neural network. This application uses a multi-scale convolutional neural network to automatically identify the weak feature background, thereby reducing manual intervention and improving recognition efficiency. The multi-scale convolutional neural network can extract multi-scale features of the background area, and the multi-scale features can accurately reflect the feature change differences of the background area, and then effectively identify the weak feature background based on the feature change differences, thereby improving the robustness and accuracy of similar background detection. It is also suitable for a variety of complex scenes and has strong generalization capabilities.

[0099] Regarding the system in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated again here.

[0100] Figure 4 It is a schematic diagram of the structure of an electronic device shown in an embodiment of the present application.

[0101] See also Figure 4 , the electronic device 400 includes a memory 410 and a processor 420.

[0102] The processor 420 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc.

[0103] The memory 410 may include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage devices. Among them, ROM can store static data or instructions required by the processor 420 or other modules of the computer. The permanent storage device may be a readable and writable storage device. The permanent storage device may be a non-volatile storage device that does not lose the stored instructions and data even after the computer is powered off. In some embodiments, the permanent storage device uses a large-capacity storage device (such as a magnetic or optical disk, flash memory) as a permanent storage device. In some other embodiments, the permanent storage device may be a removable storage device (such as a floppy disk, optical drive). The system memory may be a readable and writable storage device or a volatile readable and writable storage device, such as a dynamic random access memory. The system memory may store some or all instructions and data required by the processor at runtime. In addition, the memory 410 may include any combination of computer-readable storage media, including various types of semiconductor storage chips (such as DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), and disks and / or optical disks may also be used. In some embodiments, the memory 410 may include a readable and / or writable removable storage device, such as a laser disc (CD), a read-only digital versatile disc (such as a DVD-ROM, a double-layer DVD-ROM), a read-only Blu-ray disc, an ultra-density optical disc, a flash memory card (such as an SD card, a mini SD card, a Micro-SD card, etc.), a magnetic floppy disk, etc. Computer-readable storage media do not include carrier waves and transient electronic signals transmitted wirelessly or wired.

[0104] The memory 410 stores executable codes, and when the executable codes are processed by the processor 420 , the processor 420 can execute part or all of the methods described above.

[0105] In addition, the method according to the present application may also be implemented as a computer program or a computer program product, which includes computer program code instructions for executing some or all of the steps in the above method of the present application.

[0106] Alternatively, the present application can also be implemented as a computer-readable storage medium (or non-transitory machine-readable storage medium or machine-readable storage medium) on which executable code (or computer program or computer instruction code) is stored. When the executable code (or computer program or computer instruction code) is executed by a processor of an electronic device (or server, etc.), the processor executes part or all of the steps of the above-mentioned method according to the present application.

[0107] The embodiments of the present application have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A visual risk control method for weak feature background based on image depth analysis, characterized in that: The method comprises: Acquire the image to be detected; Inputting the image to be detected into a multi-scale convolutional neural network to extract multi-scale features of a background area from the image to be detected through the multi-scale convolutional neural network; Calculating the feature change difference of the background area based on the multi-scale features by the multi-scale convolutional neural network; The multi-scale convolutional neural network is used to identify the weak feature background in the background area according to the feature change difference.

2. The method according to claim 1, characterized in that The method further comprises: Eliminating the weak feature background from the background area to obtain a non-weak feature area; Perform similar background detection on the non-weak feature area.

3. The method according to claim 1, characterized in that The multi-scale convolutional neural network includes a first convolution kernel and a second convolution kernel of different sizes, and the multi-scale features include large-scale features and small-scale features; The extracting multi-scale features of the background area from the image to be detected by the multi-scale convolutional neural network includes: Inputting the image to be detected into the first convolution kernel and the second convolution kernel respectively through the multi-scale convolution neural network; Extracting a plurality of the large-scale features of the background area from the image to be detected by using the first convolution kernel; and A plurality of the small-scale features of the background area are extracted from the image to be detected by using the second convolution kernel.

4. The method according to claim 3, characterized in that The extracting of the plurality of large-scale features of the background area from the image to be detected by using the first convolution kernel includes: Dividing the background area in the image to be detected into a plurality of first blocks by using the first convolution kernel; Extracting feature information of each of the first blocks by using the first convolution kernel to obtain each of the large-scale features of the background area; The extracting a plurality of the small-scale features of the background area from the image to be detected by using the second convolution kernel includes: Dividing the background area in the image to be detected into a plurality of second blocks by using the second convolution kernel; Extracting feature information of each of the second blocks by using the second convolution kernel to obtain each of the small-scale features of the background area; The size of the first block is larger than that of the second block, and the feature information of the first block and the feature information of the second block both include texture, color and shape.

5. The method according to claim 4, characterized in that The calculating the feature change difference of the background area based on the multi-scale features by the multi-scale convolutional neural network includes: Calculating the feature change difference between two adjacent first blocks based on feature information of two adjacent first blocks by using the first convolution kernel; and The feature change difference between two adjacent second blocks is calculated based on the feature information of two adjacent second blocks by using the second convolution kernel.

6. The method according to claim 5, characterized in that The identifying the weak feature background in the background area according to the feature change difference by the multi-scale convolutional neural network includes: Comparing the feature change difference between two adjacent first blocks with a first change threshold through the first convolution kernel; and Compare the feature change difference between two adjacent second blocks with a second change threshold using the second convolution kernel; If the feature change difference between two adjacent first blocks is less than or equal to the first change threshold, a first recognition result output by the first convolution kernel is obtained; the first recognition result is used to characterize that the two adjacent first blocks constitute a first weak feature area; and / or, If the feature change difference between two adjacent second blocks is less than or equal to the second change threshold, a second recognition result output by the second convolution kernel is obtained; the second recognition result is used to indicate that the two adjacent second blocks constitute a second weak feature area; A sub-block commonly included in the first weak feature region and the second weak feature region is identified as a weak feature background in the background region.

7. The method according to claim 1, characterized in that The step of acquiring the image to be detected comprises: During the remote interaction with the lender, collecting visual data of the scene in which the lender is located; wherein the visual data includes at least one of image data and video data; The visual data is preprocessed to obtain an image to be detected; wherein the preprocessing operation includes at least one of image denoising and image normalization.

8. A weak feature background visual wind control system based on image depth analysis, characterized in that: The system comprises: An image acquisition module, used for acquiring an image to be detected; A feature extraction module, used for inputting the image to be detected into a multi-scale convolutional neural network, so as to extract multi-scale features of a background area from the image to be detected through the multi-scale convolutional neural network; A feature change difference calculation module, used to calculate the feature change difference of the background area based on the multi-scale feature through the multi-scale convolutional neural network; A weak feature background recognition module is used to identify the weak feature background in the background area according to the feature change difference through the multi-scale convolutional neural network.

9. An electronic device, characterized in that: include: processor; as well as A memory having executable codes stored thereon, which, when executed by the processor, causes the processor to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having executable codes stored thereon, which, when executed by a processor of an electronic device, causes the processor to execute the method according to any one of claims 1 to 7.

Images