Method for determining a confidence value of a detected object

Abstract

A method for determining a confidence value for an object of a class detected in an input image using a trained neural network is described. The method comprises the following steps: creating an activation signature for the class of the detected object using a plurality of output images of a layer of the neural network (S1), wherein the input image is supplied to an input of the neural network; the activation signature is scaled (S2) to the size of the input image; and the object portion of the scaled activation signature is compared (S3) with the distribution of activation signatures for all objects of the same class from a training dataset of the neural network to determine the confidence value. The invention also relates to a method for determining the distribution of activation signatures for a class of objects detected in a plurality of input images from a training dataset, as well as a corresponding device, a computer program, and a computer-readable storage medium.

Description

Technical Field

[0001] The invention relates to a method for determining a confidence value for an object of a class detected in an input image using a neural network. Background Art

[0002] Deep learning methods are being tested and further developed for controlling at least partially automated systems, such as autonomous vehicles or robots as examples of mobile platforms. A fundamental component of such at least partially automated systems is their ability to perceive the complexities of their surroundings. A prerequisite for the safe and efficient operation of such at least partially automated systems is the interpretation of the surrounding environment and the assessment of its state, for example, for decision-making processes such as trajectory planning and trajectory control. Deep learning methods, i.e., the use of specially trained deep neural networks, offer great potential for solving such tasks. Summary of the Invention

[0003] A fundamental safety issue with at least partially automated systems, such as those using deep neural networks, is that the ability of a neural network to solve a specific task after training can only be estimated if the input data in the application come from the same distribution as the training data set used to test and train the network. Therefore, deep neural networks (DNNs) used in object recognition tasks can only say what they see, but not how confident they are about it.

[0004] However, for example, for object recognition, knowledge about the reliability of the object recognition model is very important for at least partially automated safe driving.

[0005] Currently, within the scope of at least partially automated driving, current plausibility tests are typically limited to the comparison of multiple sensor systems or expert knowledge-based analysis of input data, for example for identifying difficult environmental conditions such as fog for optical sensors or reduced data quality such as motion blur in individual images or image sequences.

[0006] Thus, for example, if the surroundings represent This is done with the aid of a deep neural network. In the application, the data set is used to identify anomalies in the determination of a representation of the surroundings and the objects detected therein.

[0007] According to one aspect of the present invention, a method, a device, a computer program product, and a computer-readable storage medium for determining a trust value are provided, which at least partially have the aforementioned effects. Advantageous embodiments are the subject of the further technical solution and the following description.

[0008] The present invention is based on the knowledge that deviations in the detection of objects assigned to a class in an application can be identified by comparing the recognition pattern that occurs during detection with an average recognition pattern determined using a plurality of training data sets of a deep neural network. Any existing deviations can be quantified to determine a confidence value for the object detection in the application.

[0009] According to one aspect, a method is provided for determining a confidence value for an object of a class detected in an input image using a trained neural network.

[0010] In one step of the method, an activation signature for the class of a detected object is created using a plurality of output images of a hidden layer of a neural network, wherein an input image is supplied to the input of the neural network.

[0011] In a further step, the activation signature is scaled to the size of the input image. In a further step, the object portion of the scaled activation signature is compared with the activation signature distribution of all objects of the same class in a training dataset of the neural network to determine a confidence value.

[0012] Therefore, the confidence of the predictions made by a trained deep neural network (DNN) in an object recognition task can be checked by comparing the activation signature of an input image from an application scenario with the average activation signature distribution of training data from the same object category. This allows the identification of deviations between the training data and the input data during the application scenario.

[0013] In this case, the confidence value can indicate a quality number that quantifies anomalies in the detection and / or classification of an object and thus supports object recognition using a neural network by indicating additional information about the reliability of the object recognition in addition to the detection and / or classification of the object. Even if the object is correctly detected and correctly classified, such a confidence value can indicate the reliability of the method used.

[0014] Here, a trained neural network is used to describe a neural network that has been trained to solve a given task with new input data. Neural networks provide a framework for many different algorithms, such as those used for machine learning and collaboration, and for processing complex input data. These neural networks learn to perform tasks based on examples, typically without being programmed with task-specific rules.

[0015] This neural network is based on a collection of connected units or nodes called artificial neurons. Each connection can transmit a signal from one artificial neuron to another. The artificial neuron that receives the signal can process it and then activate other artificial neurons that are connected to it.

[0016] In conventional neural network implementations, the signals at the connections of artificial neurons are real numbers, and the output of an artificial neuron is calculated as a nonlinear function of the sum of its inputs. The connections between artificial neurons typically have weights that adapt as learning progresses. These weights increase or decrease the strength of the signal at the connection. Artificial neurons can have thresholds, so that they only output a signal when the total signal exceeds this threshold. Multiple artificial neurons are typically combined in layers. Different layers may perform different types of transformations on their inputs. The signal moves from the first layer, the input layer, to the last layer, the output layer, possibly after passing through these layers multiple times.

[0017] In addition to implementing a neural network, the structure of an artificial neural convolutional network (Convolutional Neural Network) consists of one or more convolutional layers, optionally followed by a pooling layer. The layer sequence can be used with or without normalization layers (e.g., batch normalization), zero padding layers, dropout layers, and activation functions (e.g., rectified linear units (ReLU), sigmoid functions, tanh functions, or softmax functions).

[0018] In principle, these units can be repeated arbitrarily often, and if repeated enough, they are called deep convolutional neural networks. Such neural convolutional networks can have a sequence of layers that scan the input grid down to a lower resolution to obtain the desired information and store redundant information.

[0019] If the data of such a neural network are described using coordinate data and feature data, wherein feature data are assigned to the coordinate data, then the amount of coordinate data typically decreases during a convolution operation, while the amount of feature data assigned to the coordinate data increases. The feature data are typically combined in so-called feature maps within the layers of the neural network.

[0020] The final convolutional layer extracts the most complex features, which are arranged in multiple feature maps (feature maps ) and generates an output image when an input image is applied to the input of the neural network. In addition, the last convolutional layer retains spatial information, which is lost in the fully connected layer that may follow, wherein the fully connected layer is used for classification.

[0021] If an input image is applied to the input of such a neural network, output images of feature maps are generated in the various hidden layers and corresponding feature maps, each of which can characterize certain features of the current input image. In a neural convolutional network, the number of feature maps is greatest in the last convolutional layer, before the coordinate dimension is converted to one dimension, for example, for classification using a fully connected layer. This makes it possible to preferably select the last convolutional layer of the neural network for creating the activation signature. The choice of layer is particularly free and can be adapted or optimized to the specific given conditions.

[0022] As convolutional layers follow one another, the feature maps become increasingly smaller in terms of their size (in terms of the number of pixels). To compare the activation signatures created based on these smaller feature maps of the convolutional layers with an input image and the objects of a class detected therein, the activation signatures are scaled to the size of the input image. Based on these scaled activation signatures of a plurality of output images (output data of the feature maps), the object parts of the classified objects detected therein are compared with the activation signature distribution of all objects of the same class in the training data set in order to determine a confidence value. For this comparison, further scaling of the object parts can be performed, in particular, since the detected objects can have different sizes in the input image.

[0023] The described scaling can also relate to a defined size or to a bounding box, respectively.

[0024] This scaling is necessary because the predicted objects usually have to be resized and normalized before the method can be applied.

[0025] In this case, the object part can be scaled with respect to the bounding box of the predicted object.

[0026] The method can be used on every pre-trained neuron convolutional network without changing or retraining it.

[0027] According to one aspect, it is proposed that the object portion of the scaled activation signature comprises a portion of the input image in which at least a portion of an object of the same category has already been detected.

[0028] In particular, the object part may comprise the entire detected object part.

[0029] According to one aspect, the object portion of the scaled output image includes a rectangular portion of the scaled output image, which is constructed around a detected object of a category.

[0030] With the aid of such a rectangular portion (English: bounding box), further image processing steps or data processing steps can be implemented more easily.

[0031] According to one aspect, it is proposed that an activation signature of a detected object of a class be determined with the aid of the following steps.

[0032] In one step of the method, the importance of each individual output image of a plurality of output images of a layer of the neural network for the classification of the detected object is calculated using a gradient method.

[0033] In a further step, each output image is weighted with its respective importance.

[0034] In a further step, the plurality of weighted output images are combined.

[0035] In a further step, an activation function is applied to the combined weighted output images in order to amplify features that have a positive influence on the classification in order to determine an activation signature.

[0036] By determining the importance of each individual output image, the subsequent weighting allows the output image of a layer that has the greatest influence on the classification of a detected object to be emphasized in the combination of multiple weighted output images. This allows us to determine which areas within the object part lead to the corresponding classification. This is an anomaly if areas outside the object part are already used for classification.

[0037] The combination of multiple weighted output images can be implemented, for example, by pixel-by-pixel addition of pixels that correspond in position.

[0038] In particular, the importance of an individual output image for the classification of a detected object can be described with a single value by averaging the importance determined by means of the gradient method over all pixels of the output image.

[0039] According to another aspect, it is proposed that an activation signature of a detected object of a class be determined with the aid of the following steps.

[0040] In one step, the importance of each individual output image of a plurality of output images of a layer of the neural network for the classification of the detected object is calculated by determining a gradient of a class output value of the neural network, which gradient relates to the output images of the plurality of output images.

[0041] In a further step, the gradients of the class output values ​​are averaged with respect to all pixels of the corresponding output image to determine the importance of each individual output image of the plurality of output images of the layer of the neural network.

[0042] In a further step, each output image is weighted according to its respective importance. In a further step, all pixel values ​​of the weighted output images that are arranged at corresponding positions in the weighted output images are summed. In a further step, the summed pixel values ​​that have a positive influence on the classification are amplified by setting all negative values ​​of the summed pixel values ​​to zero in order to determine the activation signature.

[0043] Determining the importance of individual output images of a layer using gradient methods is largely equivalent to gradient-weighted class activation mapping (Grad-CAM). It can also be performed using the Grad-CAM method. Grad-CAM is a technique for locating and visualizing class-distinguishing features in the input image, making the predictions of each convolutional neural network more transparent.

[0044] The importance of each single output image among the multiple, k output images of a layer of a neural network can be calculated with the help of Formula 1

[0045]

[0046] Here, the index c represents the category of the detected object; i and j are the indexes of the pixels of the output image, which are averaged with the help of double summation and normalization factor Z; and the category output value y c For output image The partial derivatives of .

[0047] In other words, the input image of category c and its category output value y c In the case of , the gradient (score) of the class output value with respect to the output of the feature map is calculated, where the scores of all other classes are set to zero and are globally averaged with respect to the pixels on the corresponding feature map to obtain the importance of each feature map for the object category

[0048] The activation signature can be described by the following formula 2:

[0049]

[0050] in is the activation signature of class c, with pixel indices i, j.

[0051] Here, the number k of output images with their pixel indices i, j is and importance Multiply and sum.

[0052] In other words, the output images (the output of the feature maps) are weighted by their respective importance, combined, and followed by a rectified linear unit (ReLU) to enhance features that have a positive impact on class determination. The result is a rough activation signature (heatmap) that can be superimposed on the input image to highlight important image parts.

[0053] According to one aspect, a method for determining an activation signature distribution of a class of objects detected in a plurality of input images of a training dataset is provided, wherein the training dataset contains objects of the associated class in the respective input images. The method comprises the following steps:

[0054] In one step, as described above, a plurality of activation signatures are calculated for all objects of a class detected in a plurality of input images.

[0055] In another step, each activation signature is scaled to the size of the input image.

[0056] In a further step, each object part of the object detected in the scaled activation signature is scaled to a uniform size.

[0057] In a further step, a probability distribution is calculated for each pixel of the scaled object part of the scaled activation signature using all detected objects of the same class in the training data set, wherein a mean and a variance are determined for each pixel of the scaled activation signature to determine the activation signature distribution.

[0058] According to one aspect, each pixel of the object portion of the scaled activation signature is compared with each corresponding pixel of the activation signature distribution to determine the probability that the pixel of the object portion comes from the distribution of the same pixel of the activation signature distribution, and the average of the probabilities of all pixels of the object portion is calculated to determine the trust value.

[0059] This simple method allows you to determine the trust value.

[0060] According to one aspect, it is proposed that the comparison for determining the confidence value be determined with the aid of a log-likelihood function of the scaled object portion of the activation signature and the activation signature distribution of the class of the detected object.

[0061] A different determination of the confidence value can be performed by calculation using a log-likelihood function.

[0062] According to one aspect, a comparison of the scaled activation signature of the classification object with the activation signature distribution is performed to determine the confidence value using the following relationship, which can be described by Formula 3:

[0063]

[0064] Formula 3 calculates the log-likelihood function, where x is a vector representation of the object portion of the scaled activation signature, and k is a number of pixel elements of the object portion of the scaled activation signature; μ is a vector representation of the pixel mean of the activation signature distribution with a number of k elements; and Σ is the kxk-large covariance matrix of the activation signature distribution. This means that μ and Σ are calculated using a training data set.

[0065] Here, k is the overall number of pixels, ie n times m for the x and y directions of the spatial components of the feature map, for example.

[0066] According to one aspect, a signal is generated when the confidence value is less than a confidence value threshold, wherein the threshold has been determined using training data or test data. This threshold allows, for example, an alarm signal to be generated depending on the current situation and the current confidence value. The threshold can be determined based on the confidence values ​​determined using the training and test data sets.

[0067] If, for example, a confidence value that can be determined using ln p falls below a threshold value for a new investigation, this may indicate an anomaly. A low confidence value indicates that the system is in a situation for which it has not been trained.

[0068] Furthermore, the described method can also be used as a development tool in order to already identify results that are implausible for further analysis, for example when training a neural network.

[0069] According to one aspect, a confidence value is provided as a weighting factor to the situation analysis method. Thus, a parameter is available for the situation analysis method with a confidence value, with the aid of which the result of object detection and / or classification using the neural network can be evaluated and compared with other information about the surroundings obtained using other sensors.

[0070] For example, an object detected by means of a neural network can be weighted more heavily in comparison with other sensor data when the determination is associated with a high confidence value than with a low confidence value.

[0071] According to one aspect, it is proposed that an actuation signal for actuating the at least partially automated vehicle and / or a warning signal for warning a vehicle occupant be emitted as a function of the magnitude of the confidence value.

[0072] A corresponding example of an object correctly classified as a pedestrian has been discussed above, but the location of the pedestrian in this example has a low confidence value.

[0073] A device is described which is configured to carry out the method described above. This device allows the method to be easily integrated into various systems.

[0074] A computer program is described, which includes instructions which, when the computer program is run by a computer, cause the computer to carry out one of the above-described methods. This computer program enables the described method to be used in various systems.

[0075] A computer-readable storage medium is described, on which the above-mentioned computer program is stored. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The embodiments of the present invention are Figures 1 to 3 The accompanying drawings show:

[0077] Figure 1 : Flowchart of the data flow of this method;

[0078] Figure 2 : A system having a device configured to determine a trust value;

[0079] Figure 3 : Steps of a method for determining a trust value. DETAILED DESCRIPTION

[0080] Figure 1 Describes the flow of data between components used to determine trust values.

[0081] Neural network 140 is located in the upper branch of data flow diagram 100, which depicts an application scenario. Data from various sources, such as sensors, are supplied to trained neural network 140 as input images via memory 120. This allows neural network 140 to transmit multiple output images of the hidden layer for each class of detected objects to evaluation unit 160 for each input image. Evaluation unit 160 uses the multiple output images of neural network 140 to create an activation signature and scales the activation signature to the size of the input image. Evaluation unit 160 transmits the scaled activation signature to comparison unit 180.

[0082] In the lower branch of data flow diagram 100, an activation signature is generated using a trained neural network 142. Neural networks 140, 142 have already been trained to detect and classify objects. A training data set is provided to trained neural network 142 using data memory 122, so that the neural network can generate a plurality of output images of a layer of the neural network for each of a plurality of input images.

[0083] The distribution analysis and evaluation unit 170 determines an activation signature for a category of objects detected in multiple input images; scales each activation signature to the size of the input image; scales each object part in the scaled activation signature of the detected object to a uniform size; and calculates a probability distribution for each pixel of the scaled object part of the scaled activation signature for all objects of the same category detected in the training data group to determine the activation signature distribution.

[0084] Comparison unit 180 compares the object portion of the scaled activation signature with a distribution of activation signatures from a training dataset to determine a confidence value 190 .

[0085] Figure 2 The system 200 is shown with a device 210 configured to perform a method for determining a confidence value, such as in an application. A sensor 230 provides an input image to the device 210, and an activation signature distribution is stored in a memory unit 240, so that the device 210 provides a confidence value 220 using the method described above.

[0086] Figure 3 A method for determining a confidence value is described, wherein in step S1 an activation signature for the class of a detected object is provided using a plurality of output images of a hidden layer of a neural network.

[0087] In a further step S2 , the activation signature is scaled to the size of the input image.

[0088] In a further step S3 , the scaled object portion of the activation signature is compared with the activation signature distribution of all objects of the same class in the training data set of the neural network in order to determine a confidence value.

Claims

1. A computer-implemented method for determining a confidence value for an object of a class detected in an input image by means of a trained neural network, the method comprising the following steps: Providing the input image to an input end of the neural network; An activation signature for the class of the detected object is created (S1) using a plurality of output images of a layer of the neural network; wherein, The activation signature of a detected object class is determined using the following steps: calculating the importance of each individual output image of the plurality of output images of the layer of the neural network for the classification of the detected object by means of a gradient method; Weight each output image by its respective importance; combining the weighted plurality of output images; applying an activation function to the combined plurality of weighted output images to amplify features that positively influence the classification to determine the activation signature; scaling (S2) the activation signature to the size of the input image; The object portion of the scaled activation signature is compared with the activation signature distribution of all objects of the same category in the training dataset of the neural network (S3) to determine the confidence value in order to check the credibility of the predictions of the trained neural network in object recognition.

2. The method according to claim 1, wherein The object portion of the scaled activation signature comprises a portion of the input image in which at least a portion of an object of the same class is detected.

3. The method according to claim 1 or 2, wherein: The object portion of the scaled activation signature comprises a rectangular portion of the input image constructed around a class of detected objects.

4. The method according to claim 1 or 2, wherein: The activation signature of a detected object of a class is determined with the aid of the following steps: calculating the importance of each individual output image of the plurality of output images of a layer of the neural network for the classification of the detected object by determining the gradient of the class output value of the neural network over the output images of the plurality of output images; averaging the gradients of the class output values ​​with respect to all pixels of the respective output images to determine the importance of each individual output image of the plurality of output images of the layer of the neural network; Weight each output image by its respective importance; summing all pixel values ​​of the weighted output image that are arranged at corresponding positions in the weighted output image; and The summed pixel values ​​that have a positive influence on the classification are amplified by setting all negative values ​​of the summed pixel values ​​to zero in order to determine the activation signature.

5. A method for determining an activation signature distribution of a class of objects detected in a plurality of input images of a training data set, the training data set having the associated class of the objects in the individual input images, the method comprising the following steps: Computing a plurality of activation signatures for all objects of a class detected in the plurality of input images according to claim 1 or 4; scaling each activation signature to the size of the input image; scaling each object portion of an object detected in the scaled activation signature to a uniform size; A probability distribution is calculated for each pixel of the scaled object part of the scaled activation signature using all detected objects of the same class in the training data set, wherein a mean and a variance are determined for each pixel of the scaled activation signature to determine the activation signature distribution.

6. The method according to claim 5, wherein: Each pixel of the object portion of the scaled activation signature is compared with each corresponding pixel of the activation signature distribution to determine a probability that the pixel of the object portion comes from the distribution of the same pixel of the activation signature distribution; and the average of the probabilities is calculated for all pixels of the object portion to determine the trust value.

7. The method according to claim 5 or 6, wherein: The comparison for determining the confidence value is determined by means of a log-likelihood function of the object portion of the scaled activation signature in comparison with the activation signature distribution of the class of the detected object.

8. The method according to claim 5 or 6, wherein: If the confidence value is less than a threshold value for the confidence value, a signal is generated, wherein the threshold value has been ascertained using training data or test data.

9. The method according to claim 5 or 6, wherein: The confidence value is provided as a weighting factor to the situation analysis method.

10. The method according to claim 5 or 6, wherein: Depending on the magnitude of the confidence value, an actuation signal for actuating the at least partially automated vehicle and / or a warning signal for warning a vehicle occupant is emitted.

11. A device (210) configured to perform the method according to any one of claims 1 to 10. 12 . A computer-readable storage medium having a computer program stored thereon, the computer program comprising instructions which, when the computer program is run by a computer, cause the computer to implement the method according to claim 1 .

Images