Method and system for data processing
Through the neural network identification system, the sensor data is processed in a secure environment, and the protection range of sensitive information is identified and separated, the problem of sensitive information protection of sensor data is solved, and the balance between data security and utilization is achieved.
Patent Information
- Application Number
- CN201910541299.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2018-06-22
- Filing Date
- 2019-06-21
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2039-06-21
AI Technical Summary
The prior art is difficult to effectively protect sensitive information, such as image data, measured by sensors, from loss, misuse or unauthorized access, affecting the privacy, welfare or security of an individual or organization.
Using a neural network identification system, the data measured by sensors is processed in a secure environment, and the data measured by sensors is identified and divided into parts that require the application of security features and the parts that do not require the application of security features, and the latter is output as non-safe data to ensure that Security of sensitive information.
It realizes data security when processing sensitive information, prevents unauthorized access and use, and allows further processing and use of non-sensitive data, improving the protection and utilization efficiency of information.
Smart Images

Figure CN110633576B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method and an apparatus for processing data. Background Art
[0002] Sensed data representing physical quantities measured by sensors, such as image data representing images, may be processed to detect characteristics of the sensed data. For example, image data may be processed to detect objects in the images. Sensed data (e.g., images) may contain sensitive information; the loss, misuse, or modification of which, or unauthorized access thereto, may adversely affect the privacy, welfare, or security of an individual or organization. It is therefore desirable to improve data processing methods that seek to protect such sensitive information. Summary of the invention
[0003] According to a first aspect of the present disclosure, there is provided a method for data processing for an object recognition system including a neural network, the method comprising, in a secure environment:
[0004] obtaining first sensing data, wherein the first sensing data represents a physical quantity measured by a sensor;
[0005] processing the first sensory data using the neural network to identify an object in the first sensory data;
[0006] determining that the identified object belongs to a predetermined object class;
[0007] In response to the determination, classifying a first portion of the first sensed data as data to be protected with at least one security feature and classifying a second portion of the first sensed data as data to which the at least one security feature is not applied; and
[0008] Second sensed data derived from at least the second portion is output as non-secure data.
[0009] According to a second aspect of the present disclosure, there is provided an object recognition system including a neural network, the object recognition system being configured to perform a security data processing operation, the security data processing operation comprising:
[0010] obtaining first sensing data, wherein the first sensing data represents a physical quantity measured by a sensor;
[0011] processing the first sensory data using the neural network to identify an object in the first sensory data;
[0012] determining that the identified object belongs to a predetermined object class;
[0013] responsive to the determining, classifying a first portion of the first sensed data as data to be protected with at least one security feature, and classifying a second portion of the first sensed data as data to which the at least one security feature is not applied; and
[0014] Second sensed data derived from at least the second portion is output as non-secure data.
[0015] Further features and advantages will become apparent from the following description of examples made with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 A flow chart illustrating a method of image processing for an object recognition system including a neural network, according to an example, is shown.
[0017] Figure 2 An example image processing system comprising an object recognition system for use with the image processing methods described herein is schematically illustrated.
[0018] Figure 3 A convolutional neural network according to an example is schematically shown.
[0019] Figure 4 An example neural network accelerator is schematically shown.
[0020] Figure 5 Schematically illustrating a coupled to an example deconvolutional neural network according to an example Figure 3 Convolutional neural network. DETAILED DESCRIPTION
[0021] With reference to the accompanying drawings, details of the systems and methods according to the examples will become apparent from the following description. In this specification, for the purpose of explanation, many specific details of certain examples are set forth. Reference to "example" or similar language in the specification means that a particular feature, structure or characteristic described in conjunction with the example is included in at least one example, but not necessarily in other examples. It should be further noted that certain examples are schematically described, and certain features are omitted and / or necessarily simplified for ease of illustration and understanding of the concepts underlying the examples.
[0022] Systems and methods related to processing sensed data are described herein, and the sensed data represents a physical quantity measured by a sensor. A sensor can be a device configured to measure a physical quantity (such as light or sound) and convert it into a signal (e.g., an electrical signal). Examples of sensors include image sensors and microphones. For example, an image sensor can convert light into a digital signal. The sensed data associated with an image sensor can therefore be image data representing an image captured by the sensor. The examples described below with reference to the figures relate to processing image data. However, in other examples, the sensed data can be audio data representing a sound (e.g., measured by a microphone), or can be another type of sensed data representing a different physical quantity (e.g., measured by a different type of sensor).
[0023] Figure 1 1 is a flow chart illustrating a method 100 of image processing for an object recognition system including a neural network, according to an example. An object recognition system (e.g., sometimes referred to as an object detection system or an object classification system) allows for the detection of objects belonging to a particular class of objects in an image. For example, where the class is a face, the object recognition system may be used to detect the presence of a face in an image. In some cases, the object recognition system allows for the identification of a particular instance of an object. For example, an instance may be a particular face. In this example, the object recognition system may therefore be used to identify that the detected face is a particular face that the object recognition system is arranged to identify. This may be accomplished by the object recognition system Figure 1 The object recognition system is a method 100 for image processing, such as a part of an image processing system (described below) illustratively in Figure 2 The object recognition system shown in FIG.
[0024] A neural network typically includes many interconnected nodes, which may be referred to as artificial neurons or neurons. The internal state of a neuron (sometimes referred to as the "activation" of a neuron) typically depends on the input received by the neuron. The output of the neuron may then depend on the input, weights, bias, and activation function. The outputs of some neurons may be connected to the inputs of other neurons, thereby forming a directed weighted graph in which the vertices (corresponding to neurons) or edges (corresponding to connections) of the graph are associated with weights, respectively. Neurons may be arranged in layers so that information may flow from a given neuron in one layer to one or more neurons in a continuous layer of the neural network. Examples of neural network architectures are described in more detail below.
[0025] The method involves implementing Figure 1 Items 101 to 105 shown in FIG. 101 to 105. For example, the secure environment may include a trusted execution environment. Technology to implement a secure environment on an object recognition system, the The technology provides a mechanism for implementing a security boundary in a data processing device such as an image processing system. In essence, components within a secure environment (or "security domain") are trusted within an image processing system (e.g., including an object recognition system) and are therefore allowed to access security-sensitive data, such as within the image processing system, while components outside the secure environment (e.g., in a "non-secure domain") are not allowed to access such security-sensitive data. For example, components in a secure environment may be able to access certain storage devices, such as secure or "protected" memory areas, which are not accessible to components and systems outside the secure environment. For example, components and devices that perform non-secure operations (e.g., non-secure image processing operations) should be prevented from accessing any storage devices, such as areas of memory, which are defined as secure. However, components within a secure environment (e.g., which can perform secure image processing operations) can access at least some (and in some cases all) of the secure storage devices. Certain storage devices (e.g., non-secure storage devices) may exist outside the secure environment, and components outside the secure environment (e.g., in a non-secure environment) may access them. In some cases, components within the secure environment may access such non-secure storage, for example in addition to any access to secure storage. For example, storage may be divided into a secure area and a non-secure area. Components within the secure environment may access both areas, whereas components outside the secure environment may only access the non-secure area. As described above, components such as of Such partitioning and access control can be achieved through embedded hardware technologies such as .
[0026] At item 101 of method 100, first image data associated with an image is obtained. In some cases, for example, the first image data represents the image. In other examples, the first image data is derived from image data representing the image, as described in the following examples.
[0027] Item 102 of method 100 involves processing first image data using a neural network to identify an object in the image. For example, the neural network can be configured to determine whether an object belonging to one or more predetermined object classes is present in an image corresponding to the first image data.
[0028] Item 103 involves determining that the identified object belongs to a predetermined object class. For example, the neural network can be configured to identify credit cards in an image. Thus, the predetermined object class can be "credit card," and, for example, the neural network can be configured to determine whether an object belonging to that class (i.e., credit card) is present in the image. At item 103, the neural network determines that an object belonging to a predetermined class (e.g., credit card) is present in the image.
[0029] At item 104, and in response to determining that the identified object belongs to a predetermined object class, a first portion of the first image data is classified as image data to be protected with at least one security feature, and a second portion of the first image data is classified as image data to which the at least one security feature is not applied. For example, in an example where the predetermined object class is "credit card," a first portion of the first image data corresponding to a credit card in the identified image may be classified as image data to be protected. The remainder of the image determined not to contain an object belonging to the predetermined class (e.g., credit card) may be classified as a second portion, e.g., not to be protected with at least one security feature. In some examples, there are multiple predetermined object classes. One or more of the multiple predetermined object classes may form a subset of a class, such that method 100 (at item 104) involves, in response to determining that the identified object belongs to one of the subsets of a class, classifying the first portion of the first image data as image data to be protected with at least one security feature, and classifying the second portion of the first image data as image data not to be protected with at least one security feature.
[0030] At least one security feature may be applied based on the implemented security environment. For example, applying at least one security feature to data may retain the data within the secure environment, whereas not applying at least one security feature may allow the data to be released from the secure environment. In some examples, the data to be protected (e.g., first image data) is marked (e.g., using metadata) such that only certain components with a certain access level can access the secure data. For example, a bit field or "flag" associated with the data to be protected may be set so that access to it is restricted based on components with appropriate access rights (e.g., corresponding flags). Thus, in such a case, applying at least one security feature to a first portion of the first image data may include setting a data value corresponding to a first portion of the first image data to which access is restricted. The data value may be the value of a flag or other data (e.g., metadata) corresponding to the first portion of the first image data.
[0031] As an example, a first portion of the first image data may be marked as secure by setting the value of a security flag associated therewith. For example, the associated security flag may be set to a predetermined value (e.g., to a value '1' indicating that the data is secure, rather than a value '0' indicating that the data is unsecure). Thereafter, the secure image data may be stored only in a storage device that is marked as secure accordingly, and / or may be accessed only by a component (e.g., a processor) that has access rights to the secure data (and / or secure storage device). In this example, the component may be marked as secure by also setting the value of a corresponding security flag associated with the component in a manner similar to marking the image data as secure (e.g., to a value '1' indicating that the component is secure). Thus, the first portion of the first image data may be stored only in a secure component and / or accessed by a secure component, whereas, for example, non-secure second image data (derived from at least the second portion) having a corresponding security flag value '0' may be stored in a non-secure component and / or accessed by a non-secure component. In an example, the storage and / or processing of the non-secure second image data by the secure component may be restricted. For example, the non-secure image data may not be stored in a secure storage device and / or may not be processed by a secure processor. The secure environment may implement rules, such as so that a component may only access data with the same security flag value. For example, a secure processor (e.g., having an associated security flag value of '1') may access secure data (e.g., having an associated security flag value of '1') but not non-secure data (e.g., having an associated security flag value of '0' or no security flag). As described above, in other rule examples, a secure component may access secure and non-secure storage devices or data (e.g., having corresponding associated flag values of '1' or '0'). For example, the above-referenced In some embodiments, the present invention provides a method for implementing a secure environment and / or applying at least one security feature. The method may be implemented by using a security flag value and controlling access rights in such an example. In other examples, other technologies may be used to implement a secure environment and / or apply at least one security feature. In addition, other types of security features may be applied in addition or alternatively in other examples, for example, based on the technology used to implement the secure environment.
[0032] Item 105 of method 100 involves outputting second image data derived from at least the second portion as non-secure image data. As described above, the non-secure image data may include image data without at least one security feature applied, for example, depending on the implemented security environment. For example, outputting the second image data as non-secure image data may involve writing the second image data to a non-secure storage device (e.g., a storage component to which at least one security feature is not applied and is therefore not within the secure environment). Additionally or alternatively, the second image data may be output to a non-secure domain for further processing. For example, as described above, the non-secure domain may be implemented using hardware different from the secure environment in which method 100 is implemented and / or utilizing a security protocol different from the secure environment in which method 100 is implemented. In some cases, method 100 is implemented by an object recognition system on a remote device, for example, the object recognition system uses an image sensor on the remote device to capture the first image data, and outputs the second image data from the remote device to a computing device (e.g., a server) via a network that communicatively couples the remote device and the computing device. The computing device (e.g., implemented by third-party software) may control further processing of the second image data.
[0033] In some examples, the second image data includes a second portion of the first image data. In the example above, the "credit card" class may have been set as a predetermined object class so that the portion of the image determined not to contain a credit card can be released from the object recognition system as non-secure data, for example into a non-secure domain. Other examples of predetermined object classes may correspond to other objects containing private or sensitive information, such as a child's identity document (e.g., a passport) or a human face. In some examples, the predetermined object class may correspond to an object that may include information whose authenticity will remain secure. For example, the predetermined object class may relate to a traffic sign, such as a speed limit sign, which contains traffic information to be kept safe, such as speed limit information. For example, the object recognition system may be coupled to or form part of an onboard computer device, which in such a case is coupled to or installed in a road vehicle.
[0034] In an example, the second image data includes a modified version of the first image data, wherein the first portion is obfuscated in the modified version. For example, classifying the first portion of the first image data as image data to be protected may involve, as part of the described method 100, determining a location of the identified object in the image. The first portion may then be defined based on the determined location. For example, in some examples, the first portion may include the identified object such that the entire object will be obfuscated as a portion of the first portion. In other examples, the first portion may include a portion of an object, such as a portion of a credit card including alphanumeric characters, such that only the portion of the object is obfuscated in the modified version.
[0035] In some examples, after processing the first image data using a neural network to identify an object in the image, the first image data is further processed to determine a location of the identified object in the image. In other examples, the processing for determining the location of the identified object in the image can be performed as part of or concurrently with processing the first image data to identify an object in the image.
[0036] For example, obfuscating the first portion in a modified version of the first image data may involve, for example, at least one of blurring, hiding, and redacting the first portion. Other types of obfuscation of the first portion are possible. In general, the first image data may be modified to change the appearance of the first portion in the second image data output as non-secure image data. For example, the pixel values corresponding to the first portion may be changed to obfuscate the first portion when, for example, the second image data is displayed on a display device. In this way, the first portion may be difficult to understand when the second image data is displayed. This allows the second image data to be used as non-secure data, such as for display or further processing, without releasing the original first portion of the first image data (corresponding to an object belonging to a predetermined class) from the secure environment. Thus, for example, a modified version of an image containing a credit card, wherein alphanumeric information on the credit card (such as a credit card number) is obfuscated, may be further processed outside the secure environment, for example to detect objects belonging to other classes. For example, a credit card may rest on a surface in the image, and further processing may be configured to detect such a surface in the image (e.g., for augmenting the image with a virtual object). Thus, further use of an image including a given type of object (e.g., a credit card) may be permitted outside of a secure environment without releasing specific information associated with the given type of object from the secure environment. The specific information may be sensitive, private, etc., and is intended to remain secure, e.g., not released to a non-secure domain.
[0037] In some examples, in addition to outputting the second image data including obfuscation corresponding to the first portion (e.g., the obfuscated region corresponding to the determined location of the identified object), metadata indicating that the obfuscated region is obfuscated in the second image data may also be output. In an example, the metadata may be included as part of the second image data. In other examples, the metadata may be output separately. For example, the metadata may indicate that the obfuscation in the second image data may be ignored during further processing. For example, as described, the second image data may be further processed in a non-secure domain, for example, to identify objects belonging to at least one other predetermined object class. Data (e.g., metadata) indicating that further processing may account for (e.g., ignore) the obfuscation may reduce the likelihood of errors in further processing. For example, in the case where the predetermined object class corresponds to a child's face, the second image data output by the object recognition system may include obfuscation of the image data corresponding to the location of the child's face in the image. Therefore, if the second image data is to be further processed in the non-secure domain, for example, to identify people and locate them in the image, a child with an obfuscated face may not be recognized due to the obfuscation. Thus, metadata may indicate that obfuscation (e.g., of a child's face) may be accounted for, e.g., the obfuscated region may not be processed, such that the likelihood of not identifying the child is reduced (or the likelihood of identifying the child is still increased) compared to processing the second image data without accounting for the obfuscation therein. Additionally, data (e.g., metadata) indicating that further processing may account for (e.g., ignore) the obfuscation in the second image data may reduce the number of storage accesses and / or computations for the corresponding region(s) of the second image data.
[0038] In an example, in addition to the second image data, data (e.g., metadata) may be output to indicate a location or region of interest. For example, another object belonging to another class may be detected in the image by processing the first image data. Thus, in addition to the second image data, metadata indicating a region of interest corresponding to the location of the detected other object in the image may be output. In an example where the non-secure second image data is further processed, such as by another neural network in a non-secure environment, this may allow further processing to focus on the region of interest. For example, more storage accesses and / or computations may be performed with respect to the region of interest than the rest of the image.
[0039] In some cases, as described above, it is the authenticity of information associated with a certain type of object to be kept safe. For example, returning to the example where the predetermined object class is associated with a traffic sign or in particular a speed limit sign, the first portion may correspond to alphanumeric characters on a traffic sign containing speed limit information. Therefore, image data derived from the second portion, which is classified as not protected using at least one security feature, for example because it has been determined not to contain a traffic sign identified in the image, is output as non-safe data according to item 105 of the image processing method 100. The first portion may be obfuscated and output as part of the non-safe image data, for example in a modified version of the image, as described in the example. In other examples, the first portion may not be output as part of the non-safe image data - for example, the second image data may not include image data corresponding to the first portion of the image data classified as being protected. Therefore, information associated with the identified object (e.g., speed limit information) may not be output as part of the second image data and so is not released from the secure environment. In this way, the output image data may be further used (e.g., processed) without modifying the speed limit information so that its authenticity is maintained.
[0040] In some examples, the first image data obtained, for example, at item 101 of method 100 is written to a secure storage device. For example, in addition to outputting a modified version of the image (or feature map) as non-secure image data, the original captured image (or feature data derived therefrom) may also be securely stored. Thus, the real information associated with the original image may continue to be accessible to the object recognition system or a computing device to which the object recognition system is coupled or part of. For example, the computing device may be an onboard device coupled to a road vehicle (e.g., a car) or installed in a road vehicle, and the predetermined object class may be a speed limit sign, as described in the example. Thus, the second image data output as non-secure data by the object recognition system may be further processed within the computing device, or thereby sent to an external device, and further information associated with the image may be obtained, such as an instance of an object belonging to another object class. The further information may be used by the computing device, for example, to display to a passenger of the road vehicle. The computing device may additionally or alternatively access the real speed limit information from the securely stored first image data, such as information that may have been obfuscated or not output at all as part of the first portion in the second image data. The real speed limit information can be used by the computing device, for example, to cause control of the speed of the road vehicle. In other examples, the computing device can display the real speed limit information or a corresponding warning message in the vehicle (for example, on a dashboard display).
[0041] In an example, the first image data is encrypted, for example, when being written to a secure storage device or transmitted between components within a secure environment. For example, an encryption scheme may be implemented so that the first image data is encrypted before being stored or transmitted. The first image data may be encrypted by applying an encryption algorithm to the first image data. In some examples, a hash function (e.g., a cryptographic hash function) may be applied to the first image data. Encrypting and / or hashing the first image data may ensure that the first image data has not been tampered with when it is obtained, for example, for further processing. Encryption and / or hashing may be applied to any data output from the secure environment, for example, by an object recognition system, to similarly ensure that the output data (e.g., non-secure second image data) has not been modified before being received by an external device.
[0042] Figure 2 An object recognition system 240 for use with the image processing method 100 described herein as part of an image processing system 200 is schematically shown. Figure 2 The image processing system 200 may be coupled to or form part of a computer device, such as a personal computer, laptop, smart phone, or onboard computer device that may be coupled to or installed in a vehicle such as a car.
[0043] The image processing system 200 includes an image sensor 210. The image sensor typically includes an array of sensor pixels, which can be any suitable photoelectric sensor for capturing an image. For example, a typical sensor pixel includes a photosensitive element, such as a photodiode that can convert incident light into an electronic signal or data. The sensor pixel can be, for example, a charge coupled device (CCD) or a complementary metal oxide semiconductor (CMOS). The image sensor 210 in this example is arranged to capture image data representing an image. The image can form part of a video, which is typically a series of images captured sequentially. For example, the image can correspond to a frame of a video.
[0044] exist Figure 2 In the embodiment of the present invention, the image sensor 210 is arranged to transmit image data 215 to an image signal processor (ISP) 220 of the image processing system 200 via a camera serial interface (CSI) 225. The ISP 220 can perform initial processing of the image data 215 to prepare the image data for display. For example, the ISP 220 can perform saturation correction, renormalization, white balance adjustment and / or demosaicing, but this should not be considered limiting. The ISP 220 can be secure, for example, located in a Figure 2205. Intermediate data generated by ISP 220, for example based on image data 215 transmitted thereto, and image data 230 output by the ISP may be secure. For example, components without appropriate access rights may not be able to access such secure intermediate data and / or output image data 230.
[0045] Figure 2 The object recognition system 240 is arranged to receive image data 230 from the ISP 220, which may be a processed version of the image data 215 output by the image sensor 210. However, in other examples, the object recognition system 240 may receive the image data 215 directly from the image sensor 210. For example, the object recognition system 240 may include a processor for performing initial processing of the image data 215 transmitted from the image sensor 210. In other examples, the object recognition system 240 may receive the image data 215 from a video decoder (e.g., implemented as part of a video conferencing system). For example, the image data 215 may correspond to one or more frames of a video.
[0046] As described above, the object recognition system 240 includes a neural network. The neural network can be used to process image data received by the object recognition system 240, such as the first image data 230. The object recognition system 240 can then output the second image data 245, for example, in accordance with item 105 of the method 100 described above. Figure 3 An example of a neural network 300 is schematically shown in FIG.
[0047] The neural network 300 includes a number of different layers 310, 320, 330, 340, 350. Figure 3 In FIG. 3 , the neural network 300 is a convolutional neural network (CNN). An example of a CNN is the VGG-16 CNN, but other CNNs may be used instead. A typical CNN includes an input layer 310, a plurality of convolutional layers (in Figure 3 Two of them 320, 330), multiple fully connected layers (in Figure 3 340) and an output layer 350 are shown in FIG. The input layer 310 corresponds, for example, to the input of the neural network 300, which in this example is image data.
[0048] The image data is, for example, 224 pixels wide and 224 pixels high and includes 3 color channels (such as a red channel, a green channel, and a blue channel). The convolutional layer typically extracts specific features from the input data to create a feature map and can operate on a small portion of the image. The fully connected layer then uses the feature map for classification.
[0049] exist Figure 3In the example of , the output of the second layer 320 undergoes pooling before it is input to the third layer 330. Pooling allows, for example, the values of a region or feature map of an image to be aggregated or combined, for example by taking the highest value within the region. For example, with 2×2 max pooling, the highest value of the output of the second layer 320 within a 2×2 patch of the feature map output from the second layer 320 is used as the input to the third layer 330, rather than passing the entire output of the second layer 320 to the third layer 330. This reduces the amount of computation for subsequent layers of the neural network 300. This is in Figure 3 330 as compared to the second layer 340. Further pooling is performed between the third layer 330 and the fourth layer 340 and between the fourth layer 340 and the fifth layer 350. It should be appreciated that the diagram has been greatly simplified for ease of illustration. Figure 3 The neural network 300 and typical neural networks can be significantly more complex.
[0050] Typically, such as Figure 3 Neural network 300 Such a neural network may go through what is referred to as a "training phase" in which the neural network is trained for a particular purpose. As described, a neural network typically includes many interconnected neurons forming a directed weighted graph in which the vertices (corresponding to neurons) or edges (corresponding to connections) of the graph are each associated with a weight. The weights may be adjusted throughout the training period, thereby changing the output of individual neurons and thus the output of the entire neural network as a whole. In a CNN, a fully connected layer typically connects every neuron in one layer to every neuron in another layer, and thus can be used to identify overall characteristics of an image, such as whether the image includes an object of a particular class or a particular instance belonging to that particular class.
[0051] exist Figure 3 In the example of , the neural network 300 has been trained to perform object recognition by processing image data, for example to determine whether an object of a predetermined object class is present in an image represented by the image data (although in other examples, the neural network 300 may have been trained to instead recognize other image characteristics of the image). For example, training the neural network 300 in this manner generates one or more kernels associated with at least some layers (such as layers of the neural network 300 other than the input layer 310 and the output layer 350). Thus, the output of the training may be a plurality of kernels associated with a predetermined neural network architecture (e.g., where different kernels are associated with different corresponding layers of a multi-layer neural network architecture). The kernel data may be considered to correspond to weight data representing weights to be applied to the image data, as each element of the kernel may be considered to correspond to a weight, respectively. Each of these weights may be multiplied by a corresponding pixel value of an image patch to convolve the image patch with the kernel as described below.
[0052] The kernel may allow for the identification of features of an image. For example, some of the kernels may be used to identify edges in an image represented by the image data, while other kernels may be used to identify horizontal or vertical features in the image (but this is not limiting and other kernels are possible). The precise features identified by the kernel may depend on the image features, such as object classes, that the neural network 300 is trained to detect. The kernel may have any size. As an example, each kernel may be a 3×3 matrix that may be convolved with the image data at a stride of 1. The kernel may be convolved with an image patch (or a feature map obtained by convolving the kernel with the image patch) to identify the features that the kernel is designed to detect. Convolution typically involves multiplying each pixel of an image patch (a 3×3 image patch in this example) or each element of a feature map by a weight in the kernel, and then adding the result of this operation to the result of the same operation applied to neighboring pixels or neighboring feature map elements. The stride, for example, refers to the number of pixels or feature map elements that the kernel is moved between each operation. A stride of 1 thus indicates that after the convolution is computed for a given 3×3 image patch, the kernel slides 1 pixel across the image and the convolution is computed for a subsequent image patch. This process may be repeated until the kernel has been convolved with the entire image (or the entire portion of the image to be convolved) or the entire feature map to be convolved with the kernel. The kernel may sometimes be referred to as a filter kernel or a filter. Convolution typically involves a multiplication and an addition operation (sometimes referred to as a multiply-accumulate operation). Thus, a neural network configured to implement a neural network such as Figure 4 A processor or neural network accelerator that supports a neural network may include a multiplier-accumulator (MAC) unit configured to perform these operations.
[0053] After the training phase, the neural network 300 (which may be referred to as trained neural network 300) may be used to detect the presence of objects of a predetermined object class in an input image. This process may be referred to as "classification" or "inference". Classification typically involves convolving a kernel obtained during the training phase with an image patch of the image input to the neural network 300 to generate a feature map. The feature map may then be processed using at least one fully connected layer to classify the image.
[0054] exist Figure 3In the example of , the second layer 320 involves convolution of 64 different kernels with the image data of the input layer 310. Each of the 64 kernels is, for example, arranged to recognize a different corresponding feature of the image data. In the illustrative example of image data of size 224×224 pixels, having 3 color channels and convolved with 64 kernels of size 3×3 pixels, the layer 104a of the neural network 300 involves 224×224×3×(3×3)×64 multiply-accumulate operations, i.e. 86 million multiply-accumulate operations. There will also be a large number of further multiply-accumulate operations associated with each of the third layer 330 and the fourth layer 340 of the neural network 300, which in this example involve convolution of the output of the previous layer with 256 kernels and 512 kernels respectively, where the neural network 300 is a VGG-16 convolutional neural network. However, as will be appreciated, other neural networks may involve convolution with different numbers of kernels. However, processing an image to identify image characteristics (such as the presence of a predetermined class of objects or a particular instance of an object) typically involves a large number of data processing operations, each of which consumes power.
[0055] exist Figure 3 In the example of , the image data received by the input layer 310 of the neural network 300 is processed using layers 320, 330, 340 to generate feature data. The image data may represent at least one characteristic of the light captured by the image sensor, such as the intensity of the light captured by each sensor pixel, which may be proportional to the number of photons captured by the sensor pixel. The intensity may represent the luminance of the captured light, which is, for example, a measure of the intensity of the light per unit area rather than the absolute intensity. In other examples, the image data may represent the brightness of the captured light, which may be considered to correspond to the perception of luminance, which may or may not be proportional to the luminance. In general, the image data may represent any photometric or characteristic that can be used to represent the visual appearance of the image represented by the image data, or may be derived from any such photometric or characteristic. The image data may have any suitable format, such as a raw image format. For example, the image data may be streamed from the image sensor, with or without being saved to a frame buffer, without saving the raw image data to a memory. In such a case, however, the image data obtained after processing the raw image data may be written to the memory.
[0056] In this example, the layers 320, 330, 340 of the neural network 300 may be used to generate feature data representing at least one feature of the image. The feature data may represent an output feature map, which may be obtained from, for example, Figure 3 The convolutional layer output of the neural network 300 of the CNN. In the neural network 300, there may be more than Figure 3In the example where the neural network 300 includes multiple layers 320, 330 between the input layer 310 and the fully connected layer 340 and / or the output layer 350, as shown in FIG. Figure 3 As shown in , each of the plurality of layers 320, 330 may be used to generate intermediate feature data representing at least one feature of the image. The intermediate feature data output from one of the layers (e.g., layer 320) may be input to a subsequent layer (e.g., layer 330) of the neural network 300 to identify additional features of the image represented by the image data input to the neural network 300.
[0057] Despite Figure 3 300, but it will be appreciated that further processing may be applied to the image data after it has been acquired by the image sensor and before it is processed by the layers of the neural network 300. The further processing may be performed by other components of the image processing system or as part of the neural network 300 itself.
[0058] In the example where the neural network 300 is a CNN, as in Figure 3 As in FIG. 1 , at least one layer of the neural network 300 (in this example, each of the layers 320, 330) can be a convolutional layer. The neural network 300 also includes a fully connected layer 340, which is Figure 3 The neural network 300 is arranged to receive output feature data from the convolutional layer 330, the output feature data representing, for example, an output feature map, and process the output feature data to generate object data. The object data corresponds, for example, to the output layer 350 of the neural network 300. The object data generated by the fully connected layer 340 may provide an indication of whether the neural network 300 has detected an object of a predetermined class in an image associated with the image data. Figure 3 In the example, neural network 300 includes only one fully connected layer 340. However, in other examples, neural network 300 may include multiple fully connected layers, where the output of one of the fully connected layers is received as input by a subsequent fully connected layer.
[0059] Return to Figure 2, the object recognition system 240 includes at least one processor. For example, the object recognition system 240 may include a central processing unit (CPU). However, in other examples, the object recognition system 240 may include other or alternative processors, such as a microprocessor designed to perform the functions described herein, a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any suitable combination thereof. The object recognition system 240 may also or alternatively include a processor implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors in combination with a DSP core, or any other such configuration. The object recognition system 240 may also or alternatively include at least one graphics processing unit (GPU). In the example, it may be implemented by one or more of these processors Figure 3 Neural Network 300.
[0060] In some cases, Figure 2 The object recognition system 240 may additionally or alternatively include at least one neural network accelerator. A neural network accelerator is, for example, a processor dedicated to using a neural network to at least implement classification of data. The at least one neural network accelerator may be configured according to the above-mentioned neural network examples (e.g. Figure 3 In some examples, the at least one neural network accelerator may be a neural network accelerator (such as a single or unitary neural network accelerator) configured to implement multiple neural networks (including the example neural network 300 described above). In other examples where object recognition system 240 will implement multiple neural networks, each neural network may be implemented by a corresponding neural network accelerator.
[0061] In an example, at least one processor (e.g., a neural network accelerator, a GPU, a CPU, and / or a DSP) of the object detection system 240 configured to implement the neural network 300 may support a single security domain. For example, at least one processor may remain in the secure environment 205. In such a case, the object detection system 240 may include at least one additional processor to obtain the second image data 245, for example, to perform further processing thereon. The at least one additional processor may remain outside the secure environment 205, for example, in a non-secure domain. For example, the object detection system 240 may be partitioned such that a secure partition remains within the secure environment 205 and a non-secure partition remains outside the secure environment 205. In some cases, one or more individual processors of the object detection system 240 may be segmented into respective secure and non-secure partitions. In other examples, at least one processor of the neural network 300 configured to implement the object detection system may switch between a secure domain (e.g., the secure environment 205) and a non-secure domain (e.g., outside the secure environment 205), as described in the examples herein.
[0062] exist Figure 4 The internal components of an example neural network accelerator are schematically shown in FIG. The neural network accelerator 400 includes a plurality of ALUs 410a, 410b, 410c, 410d, collectively referred to as reference numeral 410. The ALUs 410 are each arranged to perform a multiply-accumulate operation, for example, to perform the convolution discussed above. The ALUs 410 may also be arranged to each perform additional operations, such as pooling and / or activation operations.
[0063] The neural network accelerator 400 includes an interface 440 via which inputs to the neural network 400 may be received, for example, from other components of the object recognition system 240, such as a driver or storage device of another processor (e.g., a CPU). Each ALU 410 typically includes at least one MAC unit (typically an array of MAC units), but other ALUs are possible. In general, an ALU 410 may be considered to be any electronic circuit that can perform arithmetic and bitwise operations on binary numbers.
[0064] Examples of inputs that may be received by neural network accelerator 400 via interface 440 include image data representing an image to be classified using a neural network and kernel data representing kernels associated with a trained neural network. For example, the neural network may be trained by object recognition system 240 or by an external computer device or computer system (e.g., using a GPU). The output of the training may be a plurality of kernels associated with a predetermined neural network architecture (e.g., where different kernels are associated with different corresponding layers of a multi-layer neural network architecture). The kernel data may be considered to correspond to weight data representing weights to be applied to the image data, as each element of the kernel may be considered to correspond to a weight, respectively. Each of these weights may be multiplied by a corresponding pixel value of an image patch as part of a convolution of the kernel with the image patch.
[0065] Figure 4 The neural network accelerator 400 also includes direct memory access (DMA) 430, which can allow the neural network accelerator 400 to directly (e.g., independently of the CPU) access the storage device of the object recognition system 240. This can increase the speed at which the neural network accelerator 400 can perform classification of input data. For example, the neural network accelerator 400 can access input to the classification process directly from the storage device (which can be main memory or system memory) without having to wait for the CPU to be available. For example, image data, kernel data, and / or parameter data can be stored in the storage device of the object recognition system 240 and retrieved from the storage device by the neural network accelerator 400.
[0066] The neural network accelerator 400 includes a static random access memory (SRAM) 420, which can be used to store inputs to the classification process implemented by the ALU 410. For example, the SRAM 420 can be an on-chip or local memory of the neural network accelerator 400, which is generally faster to access than the storage device of the object recognition system 240. The SRAM 420 can include multiple separate storage areas. For example, image data can be stored in a different storage area than kernel data (although in some cases, they can be stored in the same storage area).
[0067] Figure 2 The object recognition system 240 including the neural network as described can perform secure image processing operations, such as those described above with reference to Figure 1One or more of items 101-105 of the described image processing method 100. In some examples, the object recognition system has a secure mode in which the object recognition system is configured to perform secure image processing operations, and the object recognition system also has a non-secure mode in which the object recognition system is configured to perform non-secure image processing operations. For example, the object recognition system 240 can be switched between a secure mode and a non-secure mode, for example, by a controller. In some cases, the secure mode and the non-secure mode of the object recognition system 240 can correspond to the secure mode and the non-secure mode of at least one processor of the object recognition system 240, respectively. Intermediate data generated by the object recognition system 240 (e.g., at least one processor) when performing secure image processing operations can be secure. For example, at least one security feature can be applied to such intermediate data, for example to limit access thereto by other components.
[0068] When in secure mode, at least one processor (e.g., a neural network accelerator) may have specific storage access restrictions. This ensures that the operation of the neural network accelerator will not cause a risk of secure data leaking from a secure domain, such as to an operating system implemented on a main processor (CPU) of an object recognition system.
[0069] In one embodiment, at least one processor as a whole (e.g., a neural network accelerator) switches between a non-secure mode and a secure mode of operation (i.e., at least one processor performs only non-secure processing or only secure processing at any one time). This avoids the need for complex security implementations within the accelerator.
[0070] In another embodiment, the switch between the secure mode and the non-secure mode is performed on an address space basis so that at least one processor can simultaneously perform both secure processing (operation) and non-secure processing (operation). In this case, a trusted firmware can be used to perform the mode switch.
[0071] The switching process can ensure that there is no state (e.g., data, etc.) in the object detection system (e.g., stored in its memory) from a secure operating mode that is maintained once the object detection system (e.g., neural network accelerator) has been switched to its non-secure operating mode. This helps reduce the possibility of data leakage from the secure domain to the non-secure domain. Therefore, in one embodiment, the process of switching the object detection system from its secure mode to its non-secure operating mode includes clearing any stored data (e.g., caches, registers, etc. storing data) from the secure operation (e.g., triggering the clearing of any stored data). The reverse switch from non-secure mode to secure mode may not involve clearing the internal state of the accelerator.
[0072] Switching between modes may be performed in any suitable manner, but in examples is performed by one (or more) of the following: using trusted firmware to control the switching; using open firmware in conjunction with trusted host software to perform the switching; entirely through trusted software; or using a hardware-implemented mechanism (e.g., where the object recognition system has a hardware internal state clearing mechanism that can be triggered when switching from a secure mode to a non-secure operating mode).
[0073] In the case where trusted firmware is used to control the switching, a (trusted) microcontroller (MCU) can be used to schedule the transition between non-secure processing and secure mode processing. For example, this can be implemented as part of a neural network accelerator (sub)system. However, other ways of controlling the switching can be used. For example, in one case a hardware mechanism is used to control the switching of an object detection system from a secure mode to a non-secure mode and vice versa. In this case, there can be a hardware mechanism that can be used to set the object detection system (e.g., a neural network accelerator) in its secure processing mode (and / or in its non-secure processing mode).
[0074] The image sensor 210 and the ISP 220 may form a secure subsystem within the secure environment 205. Thus, when the object recognition system 240 is switched into a secure mode, the secure environment 205 may be expanded to include the object recognition system. In other examples, the entire image processing system 200 (e.g., including the image sensor 210, the ISP 220, and the object recognition system 240) may be switchable between a secure mode and a non-secure (e.g., “normal”) mode. For example, as described above with reference to at least one processor of the object recognition system 240, the image sensor 210 and the ISP 220 of the image processing system 200 may be switchable between a non-secure mode and a secure mode, such as having different storage access capabilities in different modes. For example, when in their respective secure modes, the image sensor 210 and the ISP 220 may output image data to a secure storage device, and when in their respective non-secure modes, the image sensor 210 and the ISP 220 (or any other component) may not access the secure storage device. This may allow information included in an image not to be released from the secure environment 205 even before an object associated with the information has been identified in the image. Thus, information to be protected may not be released from the secure environment 205 because the non-secure second image data 245 output by the object recognition system 240 may not contain image data classified as image data to be protected (or such image data may be obfuscated). As described in the above examples, the image processing system 200 may include additional or alternative components to the image sensor 210 and the ISP 220, such as a video decoder. In some cases, components of the image processing system 200 may be switchable between a secure mode and a non-secure mode. In other cases, a component may be a secure component (e.g., permanently operating in the secure environment 205) configured to generate secure intermediate data and output secure data. Alternatively, there may be a mixture of secure and switchable components, such as an ISP 220 or a video decoder that is permanently and securely arranged to output data to an object recognition system 240 that may be switchable between a secure mode and a non-secure mode.
[0075] As previously referenced Figure 1As described in item 105 of , the object recognition system 240 may be configured to output second image data derived from at least a second portion of the first image data as non-secure image data. This output operation is one of the secure image processing operations that may be performed by the object recognition system 240. Thus, in examples where the object recognition system 240 may be switched between a secure mode and a non-secure mode, the output operation (item 105) may be performed by the object recognition system 240 when it is in the secure mode. In some examples, the object recognition system 240 may be configured to obtain the second image data and perform the non-secure image processing operation thereon when it is in the non-secure mode. For example, in addition to or in lieu of outputting the second image data when the object recognition system 240 is in the secure mode (e.g., to an external computer system for further processing), the object recognition system 240 may be switched to the non-secure mode to further process the second image data. For example, the second image data may be processed using at least one neural network of the object recognition system 240 in the non-secure mode, for example, to determine whether an image corresponding to the second image data includes an object belonging to another object class different from a predetermined object class.
[0076] In some examples, as part of outputting the second image data as non-secure image data, the second image data is transmitted to another neural network in the non-secure environment. For example, the other neural network can be part of another image processing system that is configured to receive the second image data, such as a modified version of the first image data that is obfuscated as the first portion, and process the second image data. The other neural network can have a structure that is similar to or the same as the structure of the neural network of the object recognition system. For example, the neural network and the other neural network can include a single neural network in respective secure and non-secure operating modes, as described above with reference to an object recognition system that can switch between secure and non-secure modes. In other examples, the other neural network is a separate neural network, such as that which may have been separately trained for different image processing operations.
[0077] In an example, the second image data is processed using at least one layer of another neural network to identify another object in the image. For example, the other neural network may have Figure 3The other neural network may receive the second image data at an input layer and process the second image data using one or more further layers. For example, the other neural network may process the second image data using at least one convolutional layer to generate feature data, such as one or more feature maps. The output of each layer of the other neural network may be fed as input to a corresponding subsequent layer of the other neural network. The other neural network may include at least one fully connected layer, the at least one fully connected layer being arranged to use the generated feature data, such as feature maps, for object classification (as previously described).
[0078] In some examples, the second image data output as non-secure image data (at item 105) includes feature data, such as generated by a layer of a neural network arranged to process the first image data to identify an object in the associated image (item 102). Figure 3 In the example neural network 300 of FIG. 1 , the convolutional layer 340 is arranged to receive and process the output feature data from the convolutional layer 330 to generate object data (e.g., corresponding to the output layer 350), which can provide an indication of whether the neural network 300 has detected an object of a predetermined class in an image associated with the first image data. The second image data derived from at least a second portion of the first image data (the second portion being classified as image data to which at least one security feature is not applied in response to determining that the image includes an object belonging to the predetermined object class) can include feature data generated by the intermediate layers 320, 330 of the neural network 300. As described, the second image data (e.g., including the feature data) is derived from at least a second portion of the first image data.
[0079] In some cases, one or more initial layers of the neural network 300 may correspond to one or more initial layers of another neural network, for example, the non-secure second image data may be transferred to one or more initial layers of another neural network for further processing. For example, neural networks that are trained to detect or recognize different image features or characteristics (e.g., different specific features and / or objects) still tend to have very similar or identical initial layers. For example, the initial layers of a neural network may recognize more general image features that may be present in images of a wide variety of different scenes or objects. As an illustrative example, if one neural network is trained to recognize cars in images and a different neural network is trained to recognize cats in images, the neural networks may still share similar or identical initial layers that recognize general image features such as edges, corners, or ridges in images. However, depending on what the neural networks have been trained to detect, subsequent layers in the neural networks may be different in order to perform a higher-level determination of whether a lower-level image feature corresponds to a specific object class (e.g., a car or a cat). This principle may be applied in the examples described herein to reduce processing in object recognition methods and systems. For example, the initial layers of each of the neural networks may be used to extract higher-level features that are identical to each other from the input image data.
[0080] Due to this overlap in the functionality of the initial layers of the two neural networks, one or more initial layers of the other neural network may be bypassed. For example, second image data including feature data (e.g., output feature maps (OFMs)) may be transferred from the original neural network 300 to a layer of the other neural network subsequent to its initial layer(s), e.g., without transferring the feature data to its initial layer(s). In an example, the second image data (e.g., feature data output from layer 320) is written to a (non-secure) storage device and subsequently retrieved from the storage device by a layer of the other neural network subsequent to its initial layer(s).
[0081] With this approach, rather than redoing the same initial processing as performed by one or more initial layers of the neural network, the other neural network may instead read feature data output from the initial layers of neural network 300 (e.g., from a storage device or via direct transmission of feature data), which is more efficient and may reduce the amount of processing performed by the other neural network. In some examples, one or more initial layers of the other neural network (corresponding to those of neural network 300) may be omitted when the other neural network is used for classification. For example, the other neural network may be trained so that one or more initial layers thereof correspond to one or more initial layers of neural network 300. However, the one or more initial layers may not necessarily be present in a processor or other system arranged to implement the other neural network. For example, during implementation of the other neural network for classification, such a processor or system may not include or access weight data corresponding to weights associated with the one or more initial layers.
[0082] In some examples where the second image data includes feature data, the second image data may not include feature data corresponding to the first portion. For example, feature data corresponding to the first portion that is classified as image data to be protected may not be output as part of the non-secure image data. In other examples, the second image data may include feature data corresponding to the first portion, where the first portion is obfuscated. For example, returning to the credit card example, the second image data including feature data may include an obfuscated region corresponding to an identified credit card. For example, a feature map generated during processing for identifying a credit card in an associated image may have an obfuscated region corresponding to a determined location of the credit card in the feature map. For example, obfuscation may include at least one of blurring, hiding, and redaction as previously described.
[0083] In some examples, the second image data includes feature data (e.g., a feature map) in accordance with the above examples and image data corresponding to a modified version of the image. For example, the second image data may include a feature map and a modified image, wherein the first portion is obfuscated in both the feature map and the image. A processor (e.g., a neural network accelerator) configured to process the second image data may then select the feature map or the modified image as input. For example, the second image data may be output to a plurality of processors, some of which are configured to process feature maps as input data and other processors are configured to process images as input data. Thus, a given processor may select a feature map or a modified image as input data, as appropriate for the given processor.
[0084] In the same manner that another neural network can obtain feature data output as second image data for processing—in some examples, first image data associated with an image and obtained by neural network 300 for processing may include first feature data representative of at least one feature of the image. For example, raw image data representative of an image may have been processed, for example, by a different neural network to generate the first feature data (and optionally pre-processed) before being obtained as first image data by an object recognition system. Thus, in examples the second image data may include second feature data representative of at least one feature of the image (and derived from a second portion of the first image data using neural network 300).
[0085] As previously described, classifying the first portion of the first image data as image data to be protected with at least one security feature can involve determining a location of the identified object in an image associated with the first image data. In some examples, determining the location of the identified object in the image can involve using a deconvolutional neural network to process object data output by the neural network during processing of the first image data to identify the object in the image. As described, the object data (e.g., generated by fully connected layer 340 of neural network 300) can provide an indication of whether an object of a predetermined class has been detected in the image associated with the first image data.
[0086] A deconvolutional neural network - also known as a "deconvolutional network" or a "transposed convolutional neural network" - can be considered an inverse convolutional neural network, i.e. a neural network arranged to perform a deconvolution model. Figure 5 An example of a deconvolutional neural network is schematically shown in FIG.
[0087] The deconvolution neural network 500 includes an input layer 510, for example, which corresponds to the input of the deconvolution neural network 500. In this example, the input of the deconvolution neural network 500 includes Figure 5 The object data output by the convolutional neural network 300 shown in Figure 5 The convolutional neural network 300 shown in FIG. 3 corresponds to the above-mentioned Figure 3 The corresponding description of the neural network 300 applies accordingly.
[0088] The deconvolution neural network 500 further includes a fully connected layer 520, deconvolution layers 530, 540, and an output layer 550. Figure 5 In the example of , the output of the input layer 510 undergoes unpooling before it is input to the second layer 520. Unpooling, for example, operates to substantially reverse the pooling applied by the convolutional neural network 300. For example, unpooling can allow previously aggregated or combined values for a region or feature map of an image to be ungrouped. This unpooling is performed in Figure 5Schematically shown in FIG. 5 as the size increase of the second layer 520 compared to the input layer 510. Further pooling is performed between the second layer 520 and the third layer 530 and between the third layer 530 and the fourth layer 540. It should be appreciated that the present invention has been greatly simplified for ease of illustration. Figure 5 The deconvolutional neural network 500 and typical deconvolutional neural networks can be significantly more complex.
[0089] As previously described, a deconvolutional neural network can be trained for a specific purpose. The weights associated with the neurons of the deconvolutional neural network or the connections between neurons can be adjusted throughout the training, thereby changing the output of individual neurons and thus the output of the neural network as a whole. Figure 5 In the example of , the deconvolution neural network 500 has been trained to determine the location of the recognized object in the image based on the object data obtained and associated with the image. The kernel can be deconvolved with the image data (e.g., object data, feature data) input to the layer of the deconvolution neural network 500. For example, the deconvolution layers 530, 540 are arranged to apply a deconvolution operation to the data received thereby in order to locate the object recognized by the convolution neural network 300. The output layer 550 of the deconvolution neural network 500 can correspond to the location (e.g., region) of the image or feature map that is determined (e.g., estimated) to contain the recognized object.
[0090] exist Figure 5 In the example of , convolutional neural network 300 has a first plurality of layers 310-350, and deconvolutional neural network 500 has a second plurality of layers 510-550 corresponding to the first plurality of layers. For example, at least some of the second plurality of layers 510-550 may be derivable based on corresponding layers of the first plurality of layers 310-350. For example, deconvolutional layers 530, 540 may be transposed versions of convolutional layers 330, 320, respectively. Deconvolutional neural network 500 may be considered to mirror convolutional neural network 300 in terms of its structure and / or processing. In other examples, deconvolutional neural network 500 may include more or fewer layers than convolutional neural network 300. In some examples, deconvolutional neural network 500 may have a structure that is not based on the structure of convolutional neural network 300.
[0091] In some examples, as previously described, the second image data output by the object recognition system includes feature data, such as a feature map, generated by a given convolutional layer 320, 330 of the convolutional neural network 300. In such a case, the deconvolutional neural network 500 may have layers corresponding to the given convolutional layers 320, 330 and subsequent layers 340, 350 of the convolutional neural network 300. For example, the second image data may include feature data, such as a feature map, generated by the third layer 330 of the convolutional neural network 300, and the deconvolutional neural network 500 may therefore have layers 510, 520, 530 corresponding to the third layer 330 and subsequent layers 340, 350 of the convolutional neural network 300. In an example, the deconvolutional neural network 500 may not include a layer corresponding to any layer 310, 320 of the convolutional neural network 300 before the given layer 330. Therefore, the deconvolutional neural network 500 may deconvolve the image data until the layer corresponding to the feature map transmitted as part of the second image data.
[0092] The above examples are to be understood as illustrative. Additional examples are contemplated. For example, where image data is described in the examples described with reference to the figures, for example as the first sensed data obtained, audio data representing sound may be used in other examples. For example, the first sensed data may include first audio data. In some examples, the first sensed data may include first feature data representing at least one feature of the sound. For example, a neural network may be used to process the audio data to generate a feature map representing at least one feature of the sound in a manner similar to that described for the image data. The object class when processing audio data (using a neural network) may correspond to characteristics of the audio data, such as the source of the sound, e.g., a specific person speaking or a type of vehicle making the sound. Some classes may relate to specific aspects of speech, such as the language spoken, the gender of the speaker, or the content of the speech, such as specific words or phrases.
[0093] As an example, the first sensory data may include audio data representing a spoken utterance. For example, returning to the situation where credit card information is to be kept secure, the utterance may include a credit card number. The audio data may thus be processed using a neural network to identify an object corresponding to the spoken credit card number and determine that the object belongs to a predetermined "credit card number" object class. For example, the neural network may learn the format of the credit card number as it is spoken (e.g., a string of sixteen digits spoken in groups of four) or may detect other objects that indicate an object belonging to a predetermined class (e.g., a phrase object like "My credit card number is..." is often followed by a phrase object belonging to a credit card number object class). In response to the determination, a first portion of the audio data (e.g., including the credit card number object) is classified as data to be protected with a security feature, whereas a second portion of the audio data (e.g., the remainder of the utterance) is classified as data to which a security feature is not applied. Audio data derived from at least the second portion of the input audio data is output as non-secure data.
[0094] In some examples, the audio data corresponding to the first portion to be protected may not be output as part of the output audio data. In other examples, the first portion may be obfuscated in the output audio data (as described in the reference image data example). Obfuscating the first portion in the output audio data may involve distorting the first portion of the audio data so that the corresponding audio (e.g., corresponding to the identified object belonging to a predetermined class, such as a credit card number or a specific person speaking) is difficult to understand when the output audio data is played. In an example, the first portion may be obfuscated by reducing the amplitude or volume of the first portion so that the corresponding audio is inaudible when the output audio data is played. In another example, the first portion of the audio data may be obfuscated by masking or replacing the corresponding audio with another sound (e.g., a tone or white noise).
[0095] In some examples, instead of Figure 2 The image signal processor 220 or in addition Figure 2 In addition to the image signal processor 220, another digital signal processor may be used. For example, in the case where the first sensory data includes the first audio data, an audio signal processor may be used. Similarly, references to the video decoder may be replaced with another type of decoder or codec (e.g., an audio codec in the example of processing audio data). In an alternative example where audio data is processed as the first sensory data, the image sensor 210 may be alternatively an audio sensor such as a microphone, or may be supplemented by an audio sensor such as a microphone.
[0096] Other examples are also contemplated, in which the first sensed data includes more than one type of sensed data, such as both audio data and image (or "visual") data. For example, the predetermined object class may correspond to a specific person. Audio data and visual data (e.g., image data, video data) may be processed into the first sensed data to identify a specific person therein. For example, a person may be identified as appearing in an image of the first sensed data, and / or speaking in the audio data of the first sensed data. The first portion of data classified as data to be protected may therefore include a corresponding first portion of audio data and / or a corresponding first portion of image data. Therefore, in an example in which the first portion is obfuscated in the non-secure output data, the audio data and / or image data may include the obfuscated first portion. For example, a specific person may be obfuscated, such as removed or masked, in the output image data and / or the output audio data. In another example, a specific person may be blurred in the output image data (e.g., making the person difficult to identify), and a voice corresponding to the person may be distorted in the output audio data (e.g., making the person difficult to identify).
[0097] It should be understood that any feature described with respect to any one example may be used alone or in combination with other features described, and may also be used in combination with one or more features of any other example or any combination of any other example. In addition, equivalents and modifications not described above may also be adopted without departing from the scope of the appended claims.
Claims
1. A method for data processing for an object recognition system, the object recognition system comprising a neural network, the method comprising performing the following operations in a secure environment in the object recognition system: obtaining first sensing data, wherein the first sensing data represents a physical quantity measured by a sensor; processing the first sensory data using the neural network to identify an object in the first sensory data; determining that the identified object belongs to a predetermined object class; responsive to the determination, classifying a first portion of the first sensed data as data to be protected with at least one security feature and classifying a second portion of the first sensed data as data to which the at least one security feature is not applied; as well as Second sensed data derived from the first portion of the first sensed data and the second portion of the first sensed data is output as non-secure data for processing outside the secure environment, wherein data corresponding to the first portion of the first sensed data is obfuscated in the second sensed data.
2. The method according to claim 1, wherein: The first sensing data includes: first image data representing an image; first feature data representing at least one feature of the image; first audio data representing a sound; or First feature data representing at least one feature of the sound.
3. The method according to claim 1, wherein: Outputting the second sensing data includes: writing the second sensing data to a non-secure storage device; and / or transmitting the second sensed data to another neural network in a non-secure environment, Outputting the second sensed data optionally further comprises: processing the second sensed data using a layer of the another neural network to identify another object in the sensed data, Optionally, the neural network and the another neural network comprise a single neural network in a safe operating mode and a non-safe operating mode, respectively. The method of claim 1 , comprising writing the first sensed data to a secure storage device.
5. The method according to claim 1, wherein: Classifying the first portion of the first sensed data as data to be protected includes: determining a location of the identified object in the first sensory data; and The first portion is defined based on the determined location.
6. The method according to claim 5, wherein: Determining the location of the identified object in the first sensory data includes: processing object data using a deconvolutional neural network, the object data being output by the neural network based on processing the first sensory data to identify the object, Optionally, the neural network includes a convolutional neural network having a first plurality of layers, and the deconvolutional neural network includes a second plurality of layers corresponding to the first plurality of layers.
7. A method according to any one of the preceding claims, wherein: The second sensing data: Included is second feature data derived from the second portion of the first sensory data using the neural network.
8. An object recognition system, the object recognition system comprising a neural network implemented in a secure environment in the object recognition system, the object recognition system being configured to perform secure data processing operations, the secure data processing operations comprising: obtaining first sensing data, wherein the first sensing data represents a physical quantity measured by a sensor; processing the first sensory data using the neural network to identify an object in the first sensory data; determining that the identified object belongs to a predetermined object class; responsive to the determination, classifying a first portion of the first sensed data as data to be protected with at least one security feature and classifying a second portion of the first sensed data as data to which the at least one security feature is not applied; as well as Second sensed data derived from the first portion of the first sensed data and the second portion of the first sensed data is output as non-secure data for processing outside the secure environment, wherein data corresponding to the first portion of the first sensed data is obfuscated in the second sensed data.
9. The object recognition system according to claim 8, wherein: The object recognition system comprises: a secure mode in which the object recognition system is configured to perform the secure data processing operations; and A non-secure mode in which the object recognition system is configured to perform non-secure data processing operations.
10. A data processing system, comprising: The object recognition system according to claim 8 or 9; as well as Components configured to perform secure data processing operations; Therein, the object identification system is arranged to receive the first sensed data from the component.
Citation Information
Patent Citations
Systems and methods for generating accurate sensor corrections based on video input
CN105103089A
Computer-vision-based security detection method
CN106156765A
Gender recognition method and device based on multi-output convolution neural network, and computing equipment
CN106295521A
Devices, methods and systems for biometric user recognition utilizing neural networks
CN107533642A
Vehicle Ambient Audio Classification Via Neural Network Machine Learning
CN107967917A
