Method for configuring an object recognition system

Abstract

A method for configuring an object recognition system (100) comprises the following steps: providing labeled training data (3), the labeled training data (3) comprising image data (1), the image data (1) having a defined assignment relationship (2) to at least one object; training a neural network (40) having a first neural subnetwork (20), the first neural subnetwork (20) being configured to perform compression of the image data (3), wherein the first neural subnetwork (20) is interconnected with at least one further neural subnetwork (30a, ..., 30n), wherein at least one further neural subnetwork (30a, ..., 30n) is configured to recognize an object from the compressed training data (3); parameterizing the first neural subnetwork (20) such that object recognition is performed with the aid of the at least one further subnetwork (30a, ..., 30n) with a defined quality; and jointly training the neural subnetworks (20, 30a, ..., 30n).

Description

Technical Field

[0001] The present invention relates to a method for configuring an object recognition system. Furthermore, the present invention relates to a method for identifying an object using an object recognition system. Furthermore, the present invention relates to a sensor device. Furthermore, the present invention relates to an object recognition system. Furthermore, the present invention relates to a computer program. Furthermore, the present invention relates to a machine-readable storage medium. Background Art

[0002] Known data compression methods, for example based on discrete cosine transforms or wavelet transforms, are frequently used in sensor systems to compress raw data, with the compressed raw data typically being transmitted via a data bus for further processing. This compression is necessary in multi-sensor systems because the data bus is a major bottleneck, and the aforementioned data compression can reduce the amount of data or the data rate via the data bus.

[0003] Known methods for compressing sensor data generally minimize common information-theoretic metrics. However, compression is not biased against further processing of the data using machine learning techniques, which later interpret the data. Consequently, suboptimal performance may not be achieved during further processing, as important data may be lost during compression or the optimal compression ratio may not be achieved.

[0004] It is known to perform data compression on the sensor side using standard methods (eg JPEG for image processing) in order to reduce the amount of transmitted data. Summary of the Invention

[0005] The object of the present invention is to provide a method for identifying objects in an improved manner by means of an object recognition system.

[0006] According to a first aspect, the object is achieved by a method for optimizing an object recognition system, the method comprising the steps of:

[0007] - providing labeled training data, which include image data having a defined assignment to at least one object;

[0008] - training a neural network having a first neural subnetwork, which is configured to perform compression of image data, wherein the first neural subnetwork is interconnected with at least one other neural subnetwork, wherein

[0009] - at least one other neural sub-network configured to recognize objects from the compressed training data;

[0010] wherein a first neural sub-network is parameterized such that object recognition is performed with a defined quality by means of at least one further sub-network; and wherein the neural sub-networks are trained jointly.

[0011] In this way, during the training phase, the neural subnetwork is configured such that it is suitable for use in an object recognition system. The degree of compression of the image data is advantageously adapted to the object recognition to be achieved, thereby making it possible to find a favorable compromise between the two objectives (data compression and object recognition quality) using this method.

[0012] Advantageously, data compression of an object recognition system can be performed during operation so that, for example, unimportant sensor data can be identified and thus compressed more strongly. In contrast, more important sensor data is compressed less, so that object recognition is still sufficiently good with the reduced data volume. The proposed training method is performed offline before the object recognition system is operated using the optimized compression mechanism. Advantageously, the data bus to which one or more sensor devices are connected can be better utilized in this way, or more sensor devices can be connected to the data bus. As a result, data compression is "set" for the target application, which supports high-quality object recognition for the target application while simultaneously achieving high data compression.

[0013] According to a second aspect, the object is achieved by a method for identifying an object by means of an object recognition system, the method comprising the steps of:

[0014] - compressing the image data by means of a compression device trained on the first neural sub-network; and

[0015] - Object recognition from the compressed image data is performed with the aid of an evaluation device that has been trained on the basis of at least one further neural sub-network.

[0016] According to a third aspect, the object is achieved with an object recognition system having a defined number of sensor devices functionally connected to one another and a computer device functionally connected to the sensor devices for performing object recognition using compressed sensor data.

[0017] According to a fourth aspect, the object is achieved with a sensor device having a compression device configured by means of a method for configuring an object recognition system.

[0018] According to a fifth aspect, the object is achieved by means of an object recognition system having a defined number of the proposed sensor devices which are functionally connected to one another, and having a computing device which is functionally connected to the sensor devices via a data bus and is used to perform object recognition from compressed sensor data.

[0019] According to a sixth aspect, the object is achieved by a computer program comprising instructions which, when the computer program is executed by a computer, cause the computer to carry out a method for configuring an object recognition system.

[0020] According to a seventh aspect, the object is achieved by a computer program comprising instructions which, when the computer program is executed by a computer, cause the computer to carry out a method for recognizing an object by means of an object recognition system.

[0021] According to an eighth aspect, the object is achieved by means of a machine-readable storage medium on which at least one of the aforementioned computer programs is stored.

[0022] Preferred embodiments of the method and the object recognition system are the subject matter of the dependent claims.

[0023] An advantageous embodiment of the method is characterized in that the parameters of the first neural subnetwork define the degree of compression of the training data, while the parameters of at least one further neural subnetwork define the quality of object recognition. In this way, a compromise is achieved between the degree of compression using the first neural subnetwork and the quality of object recognition using the at least one further neural subnetwork.

[0024] Another advantageous embodiment of the method is characterized in that the degree of compression of the training data is defined by a first objective function, wherein the object recognition quality is defined by at least one further objective function. This advantageously allows the properties of the neural subnetworks or the goals to be achieved with them to be precisely defined.

[0025] Another advantageous embodiment of the method is characterized in that the weights of the first neural subnetwork are determined by training the first neural subnetwork, wherein the weights of the first neural subnetwork and the weights of at least one further neural subnetwork are determined by training the at least one further neural subnetwork. Advantageously, this provides a higher-level general training scenario in which the weights of the first neural subnetwork are also trained by training the further neural subnetworks.

[0026] Another advantageous embodiment of the method is that regions of training data with high entropy are compressed less than regions of training data with low entropy, wherein low entropy training data include, for example, image data with uniform colors and / or image data with recurring patterns. This allows for typical application scenarios such as compressing image data and performing associated object recognition from these image data. The lower the entropy of the training data, the easier it is to compress such data.

[0027] Another advantageous embodiment of the method is characterized in that the second objective function is defined from at least one of the following: identification of a person, identification of a vehicle, and identification of infrastructure. This allows for typical application scenarios for object recognition from compressed image data.

[0028] A further advantageous embodiment of the method is characterized in that simplified image data with reduced entropy are generated from the image data and that these simplified image data are subsequently compressed by means of a parameterized standard compression device.

[0029] Another advantageous embodiment of the method is characterized in that the image data is compressed using a parameterized standard compression device. In this case, the image data is compressed using a parameterizable standard compression device. Advantageously, a standard compression device (e.g., JPEG compression for image data) can also be used in a manner specific to object recognition systems.

[0030] In the following, the invention is described in detail with further features and advantages with reference to a number of figures. Identical or functionally identical elements are provided with the same reference numerals.

[0031] The disclosed method features are derived analogously from the corresponding disclosed device features, and vice versa. This means, in particular, that the features, technical advantages, and embodiments of the proposed method are derived analogously from the corresponding embodiments, features, and advantages of the sensor device or object recognition system, and vice versa. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In the attached figure:

[0033] Figure 1 shows a schematic diagram of the proposed object recognition system in training mode;

[0034] Figure 2 A schematic diagram showing the proposed object recognition system in normal operation at run time;

[0035] Figure 3A schematic diagram showing a variant of the proposed object recognition system during normal operation at runtime; and

[0036] Figure 4 A schematic diagram of an embodiment of the proposed method is shown. DETAILED DESCRIPTION

[0037] Various types of object recognition are known, such as object detection and semantic segmentation. The term "object recognition" is generally understood to mean object detection, although semantic segmentation is also a form of object recognition. Object recognition is also understood to be a machine learning method that interprets sensor data for the purpose of understanding the surrounding environment.

[0038] For example, in object recognition, such as pedestrian detection, open areas of the image are of little or no importance for the application and can therefore be compressed more strongly. However, these areas may also have high entropy or texture (e.g., a cloud-covered sky, complex road texture, etc.), so that they are not optimally compressed using conventional standard compression methods.

[0039] This is advantageously supported by the proposed image-adaptive, trained or optimized data compression, with knowledge of the “object recognition quality” of the target application.

[0040] The proposed method is based on multi-objective optimization and on a previously trained or optimized compression device. For optimization purposes, for example, multiple neural sub-networks can be connected in series or in parallel and trained jointly.

[0041] Figure 1 A schematic diagram of a training scenario for object recognition system 100 is shown. Labeled training data 3 is shown, comprising image data 1 and labels 2, wherein the labels 2 include defined assignments of objects within image data 1 (e.g., coordinates of a person in the image data). The assignment of labels 2 to image data 1 occurs in a manner known per se through human activity.

[0042] Prior to the training mode of the object recognition system, labeled image data 3 are provided, that is to say a large number of manually labeled images, thereby providing specific information for the subsequent training process.

[0043] As can be seen, the image data 1 are fed to a first neural subnetwork 20 of a neural network 40, which is functionally interconnected with at least one further second subnetwork 30a, ..., 30n. The first neural subnetwork 20 is defined by a first target function Z1, which specifies the degree of data compression of the image data 1. The further neural subnetworks 30a, ..., 30n are each defined by further target functions Z1. 2a ,...,Z 2n To define, the other objective function Z 2a ,...,Z 2n The quality of object recognition from the compressed image data 1 is respectively defined.

[0044] "Object recognition quality" is understood here to mean the deviation between the labeled image data and the object recognition performed by the other neural subnetworks ("object recognition error"). Consequently, the quality of object recognition can be expressed as the error between human annotation and machine object recognition.

[0045] In accordance with Figure 1 The training time of the training scenario of the individual neural sub-networks 20, 30a, ..., 30n is combined into the overall neural network 40 and trained jointly. The objective function of the overall network 40 is the objective function Z1, Z2 of the individual neural sub-networks. 2a ,...,Z 2n The first neural subnetwork 20 determines an intermediate representation of the image data 1. The quality of the intermediate representation ("compressibility") is evaluated using a first objective function Z1 ("compression objective function"). This intermediate representation then serves as input for one or more further neural networks 30a, ..., 30n ("application network"), which in turn calculate their outputs from this representation. The quality of the respective output is evaluated using further objective functions Z1 and Z2. 2a ,...,Z 2n In these objective functions, the deviation of the network output from the application-specific annotations is calculated (e.g., the location of pedestrians in the image, the location of infrastructure in the image, etc.).

[0046] Using the objective functions Z1, ..., Z n To strive for, while optimizing other objective functions Z 2a ,...,Z 2n In the case of , the first objective function Z1 is optimized. The first objective function Z1 preferably defines the degree of data compression or the compression rate, so that the compressed image data or sensor data can be transmitted via a data bus (not shown) at the lowest possible data rate.

[0047] The labeled image data 3 are compared with the outputs of the other neural sub-networks 30a, ..., 30n. The resulting error yields a gradient, which is used in a gradient descent method to adapt the gradients or weights of the first neural sub-network 20 so that the other neural sub-networks 30a, ..., 30n achieve the most reliable possible object recognition, for example, in the form of the position of a person in the image.

[0048] The other neural sub-networks 30a, ..., 30n are initially initialized with random weights, which leads to errors in object recognition. The goal is for the other neural sub-networks 30a, ..., 30n to recognize objects as they exist in the form of labeled image data 3. This allows for the most accurate and reliable object recognition possible using compressed image data 1.

[0049] In a variant of the training process, standard compression devices or algorithms 11a, ..., 11n can be used for training, wherein high-resolution image data 1 is fed to the standard compression devices 11a, ..., 11n. In a first sub-variant, simpler image data are first generated from the high-resolution image data 1, wherein entropy can be reduced, for example, by identifying defined image structures in the image data 1. The simplified image data 1 are then compressed using the standard compression devices 11a, ..., 11n. These images can then be used in decompressed form to train the first neural sub-network 20.

[0050] In a second sub-variant, the standard compression devices 11a, ..., 11n use the predicted parameters. High-resolution raw image data 1 is supplied to the standard compression devices 11a, ..., 11n along with the predicted parameters, with the data compression being controlled by these parameters. Parameters optimized for the current image content can then be determined or estimated for the standard compression devices 11a, ..., 11n. In this case, the image data is also subsequently decompressed and can be used to train the first neural sub-network 20.

[0051] According to the trend, in this variant, although the object recognition results are less accurate, the standard compression algorithm can be used. As a result, after the training process is completed, the trained standard compression algorithm is provided together with the other trained neural networks 30a, ..., 30n.

[0052] Since the neural sub-networks 20, 30a, ..., 30n are jointly trained as a common neural network 40, each free parameter in all sub-networks 20, 30a, ..., 30n is subject to all objective functions Z1, Z2, Z3 and Z4. 2a ,...,Z 2nThis means, in particular, that the parameters in the first subnetwork 20 are not only selected to achieve the best possible compressed intermediate representation, but also to retain the signals that are crucial for the application network 30a, ..., 30n in each case. In this way, the first neural subnetwork 20 ("compression network") is informed of which signals are important for the application and must not be compressed, and which signals are unimportant and can therefore be compressed strongly or discarded.

[0053] For example, it is conceivable that image regions or pixels with the same color are compressed more strongly than image regions or pixels with different colors. Furthermore, it is also conceivable that image regions of image data 1 with a high amount of texture or entropy should retain as much of their information content as possible and thus be compressed only slightly compared to image regions of image data 1 with a low amount of texture or entropy.

[0054] As a result, after the training process has concluded, the weights of the neural sub-networks 20 , 30 a , . . . , 30 n are determined, which are then subsequently used as trained algorithms in the object recognition system 100 during normal operation.

[0055] The first objective function Z1 evaluates the compressed sensor data, and the other objective functions Z 2a ,...,Z 2n The first and further neural networks 20, 30a, ..., 30n are affected. As a result, the neural network 40 is not trained or optimized in an "application-independent" manner, but rather in an "application-aware" manner.

[0056] Therefore, in Figure 1 In the configuration or training scenario of , it is proposed that the compression device in the form of a first neural subnetwork 20 and the evaluation devices in the form of further neural subnetworks 30a, ..., 30n are connected together to form an overall system in the form of a neural network 40 and are jointly optimized. As the objective function, a combined objective function consisting of an objective function for compression and an application-specific objective function is used. The first objective function Z1 evaluates the quality of the compressed sensor data, while the further objective functions Z 2a ,...,Z 2n Evaluate the final object recognition using the compressed sensor data.

[0057] exist Figure 2, an embodiment of the proposed object recognition system 100 is schematically shown. A plurality of sensor devices 10a, ..., 10n (e.g., LiDAR sensors, radar sensors, ultrasonic sensors, cameras, etc.) can be seen, which are installed, for example, in a vehicle and are configured to determine image data 1 when detecting the vehicle's surroundings. In each of these sensor devices 10a, ..., 10n, a compression device 11a, ..., 11n can be seen. Each of these compression devices 11a, ..., 11n comprises a first neural subnetwork 20 trained according to the present invention or a standard compression device trained according to the present invention, with which the image data 1 detected by the sensor devices 10a, ..., 10n is compressed into compressed sensor data S. ka ,...,S kn And it is supplied to the data bus 4 (for example, a CAN bus).

[0058] Via data bus 4, the compressed sensor data S ka ,...,S kn is supplied to a central processing unit 50 in which at least one evaluation device 31a, ..., 31n is arranged. ka ,...,S kn , the evaluation devices 31 a , . . . , 31 n perform object recognition (eg pedestrian recognition, vehicle recognition, traffic sign recognition, etc.) based on the other neural sub-networks 30 a , . . . , 30 n .

[0059] In this way, an object recognition system 100 is provided, which has one or more sensor devices 10a, ..., 10n, each of which has a sensor-side computing unit in the form of a compression device 11a, ..., 11n for data compression, and one or more central computing units 50 for further processing the compressed image data. The central computing unit 50 generally has a computing power that is several orders of magnitude greater than the sensor-side computing unit or the first neural subnetwork 20 of the sensor devices 10a, ..., 10n.

[0060] The proposed object recognition system 100 can reduce the amount of data transmitted from sensor devices 10a, ..., 10n to central processing unit 50 without adversely affecting the functional performance of object recognition system 100. To this end, image data 1 is compressed on the sensor side using parameterized or trained compression devices 11a, ..., 11n. The algorithms of parameterized compression devices 11a, ..., 11n can be trained, for example, using first neural subnetwork 20 in the manner described above; in an alternative embodiment, compression devices 11a, ..., 11n can also be designed as parameterizable standard compression methods.

[0061] In this case, the neural network on the sender side modifies this data beforehand so that standard methods can achieve better compression results. Standard compression methods also have to strike a balance between the following two objectives:

[0062] - Compression quality: how strongly the data are compressed according to the first objective function Z1

[0063] - Compression quality: how well the original data is preserved through compression, or how large the error is caused by compression

[0064] For other objective functions Z 2a ,...,Z 2n In standard compression methods, common metrics (such as entropy) are used. However, in this case, no semantic knowledge about the scene is incorporated. As explained above, certain image regions may have high entropy, but they are not necessarily important from the perspective of object recognition. Despite high entropy, it is entirely possible that these image regions have certain characteristics (such as typical color gradients, typical edge structures, etc.) that "meaninglessly" exclude the object being searched for.

[0065] Here, the proposed approach starts by replacing the common metric used to measure compression quality with a metric that obtains semantic knowledge of the target application (e.g. object recognition). While entropy can be expressed as a closed objective function, the new metric is defined by the following factors:

[0066] - Objective function for object recognition

[0067] - A model used to perform object recognition. In the case of a neural network, this is the network structure and its free, "learnable" parameters

[0068] - 3 labeled training data, which is used to learn object recognition

[0069] While entropy is a fixed measure, the compression measure varies during the joint training process of the neural network 40, since the object recognition itself even has learnable parameters. Thus, the "compression network" 20 learns which image regions / texture patterns can be simplified / compressed without sacrificing the object recognition performed by the "application network" 30a, ..., 30n, and simultaneously adapts the object recognition to the learned compression.

[0070] This results in better results than when the components are considered in isolation, as is the case with standard compression methods. The compressed image data S is transmitted to the central processing unit 50. ka ,...,S kn Optionally, the compressed image data S ka ,...,S kn There, the decompressed or compressed image data is used for object recognition. Finally, the compressed image data is further processed by one or more evaluation devices 31a, ..., 31n (eg, object detection and / or semantic segmentation using the compressed sensor data).

[0071] therefore, Figure 2 The object recognition system 100 shows a diagram of the neural network 40 after a completed training phase. The trained neural sub-networks 20, 30a, ..., 30n are now used in the sensor devices 10a, ..., 10n or the central processing unit 50, wherein the neural sub-networks 20, 30a, ..., 30n are no longer required to be learned during the runtime of the object recognition system 100. For this purpose, a new training process may have to be started using new labeled training data 3.

[0072] Other objective functions Z 2a ,...,Z 2n The optimization may consist, for example, in that an object (e.g. a pedestrian, a vehicle, an object, etc.) is recognized as well as possible by one of the sensor devices 10a, ..., 10n. In this way, the objective functions Z1, ..., Z2 of the neural subnetworks 20, 30a, ..., 30n are thus optimized collectively and simultaneously. 2n .

[0073] As a result, the optimized, compressed image data S is used. ka Adequate or optimized object recognition can also be achieved.

[0074] Advantageously, by using machine learning techniques in a coordinated manner both in compression and in reprocessing, the proposed method is able to either achieve a higher compression rate than would be possible with conventional compression or to achieve better results in reprocessing (e.g. object recognition) than would be possible with conventional compression.

[0075] As you can see, in Figure 2 In the example, a plurality of evaluation devices 31a, ..., 31n can be provided within the computing device 50, wherein in this case the same compressed image data S k To supply all evaluation devices 31a, ..., 31n. In a simple manner, multiple optimizations are thereby possible.

[0076] Compression can also be achieved without an explicit target function by selecting the number of output neurons of the compression network to be smaller than the dimensionality of the image data 1. In this way, the first target function can be provided solely by the structuring of the compression device 11a (not shown in these figures).

[0077] The application-specific objective function may, for its part, comprise a plurality of objective functions Z1, . . . , Z n combination.

[0078] The core idea of the proposed method is therefore to provide compression for subsequent reprocessing of the compressed sensor data.

[0079] For example, a signal processing chain might look like this:

[0080] - In the sensor, an image with assigned image data 1 is recorded

[0081] - On the sensor side, the signal is compressed using a "compression network" 20

[0082] - Send compressed image data to the central computing device 50 via the data bus 4

[0083] In the central processing unit 50 , the compressed image data are processed via the application-specific neural network.

[0084] example:

[0085] The input image is 640×480 pixels in size, with three input channels (RGB—red, green, and blue). The number of input neurons = the number of R, G, and B pixels = 640×480×3 = 921,600 neurons. The output layer for the intermediate representation is constructed as a grid of 320×240 neurons (76,800 neurons), which means that the size of the data to be transmitted via the data bus is reduced by a factor of 12.

[0086] In other words, this means that instead of an RGB image with a size of 640×480, only a single-channel image with a size of 320×240 must be sent via the data bus.

[0087] Figure 3 A further embodiment of the proposed object recognition system 100 is schematically shown. In this case, in the sensor device 10a, the image data 1 are compressed by means of a first neural sub-network 20 and subsequently fed to the encoder 5. At the output of the encoder 5, the compressed image data S K is sent to the data bus 4. From the data bus 4, the compressed image data S K The data are fed to a decoder 6 which distributes the data to the neural sub-networks 30a, ..., 30n. Even in this variant, the compression and object recognition are divided between the sensor device 10a and the central processing unit 50.

[0088] The "compression network" calculates an intermediate representation that is smaller, of the same size, or larger than the original sensor or image data. However, this intermediate representation is manipulated so that it can be effectively and losslessly compressed using other methods. On the sensor side, a compression unit is installed that implements standard methods for lossless compression (e.g., entropy coding). A corresponding decompression unit in the form of a decoder 6 is provided on the central processing unit 50. In this case, the signal processing chain looks like this:

[0089] - In the sensor, an image with the assigned image data is recorded

[0090] - Compute an intermediate representation by compressing the network. This step is generally lossy, that is, the original data cannot be reconstructed without error from the intermediate representation

[0091] - Lossless compression of intermediate representations on the sensor side

[0092] - Send the compressed intermediate representation to the central computing device 50 via the data bus 4

[0093] - Decoding of the compressed intermediate representation by a decompression unit in the form of a decoder 6

[0094] - reprocessing the decompressed intermediate representation through the / the application-specific neural network.

[0095] During the training phase, the compression and decompression themselves do not have to be transformed, that is, the graph of the training process does not change.

[0096] example:

[0097] Input image of size 640 × 480 pixels, 8 bits per pixel

[0098] Intermediate Representation: The processed image of size 640×480 with 8 bits per pixel. However, by compressing the network, the entropy of the data has been reduced.

[0099] Compression can now be performed using standard methods such as, for example, entropy coding.

[0100] Figure 4 A basic flow chart of an embodiment of the proposed method is shown.

[0101] In step 200 , labeled training data 3 are provided, which include image data 1 with a defined assignment 2 to at least one object.

[0102] In step 210, a neural network 40 having a first neural sub-network 20 is trained, which is configured to perform compression of training data 3, wherein the first neural sub-network 20 is interconnected with at least one other neural sub-network 30a, ..., 30n, wherein the at least one other neural sub-network 30a, ..., 30n is configured to recognize objects from the compressed training data 3.

[0103] In step 220 , the first neural subnetwork 20 is parameterized such that object recognition is performed with a defined quality using at least one further subnetwork 30 a , . . . , 30 n , wherein the neural subnetworks 20 , 30 a , . . . , 30 n are trained jointly.

[0104] The proposed method is preferably designed as a computer program having program code means for carrying out the method on a compression device and on an evaluation device. Both the compression device and the evaluation device can be implemented in software, wherein the compression device is copied to the sensor device after optimization and is set up there for normal operation.

[0105] Although the present invention has been described above based on specific embodiments, those skilled in the art may implement embodiments that are not disclosed above or are only partially disclosed without departing from the core of the present invention.

Claims

1. A method for configuring an object recognition system (100), comprising the steps of: - providing labeled training data (3), said labeled training data (3) comprising image data (1), said image data (1) having a defined assignment relationship (2) to at least one object; - training a neural network (40) having a first neural sub-network (20), the first neural sub-network (20) being configured to perform compression of the training data (3), wherein the first neural sub-network (20) is interconnected with at least one other neural sub-network (30a, . . . , 30n), wherein - the at least one further neural sub-network (30a, . . . , 30n) is configured to recognize objects from the compressed training data (3); wherein the first neural sub-network (20) is parameterized so that object recognition is performed with a defined quality by means of the at least one further neural sub-network (30a, . . . , 30n); And among them, The first neural sub-network (20) and the at least one other neural sub-network (30a, ..., 30n) are jointly trained by: 1) evaluating a first objective function (Z1) based on the output of the first neural sub-network (20), wherein the degree of compression of the training data (3) by the first neural sub-network (20) is defined by the first objective function (Z1), 2) evaluating at least one other objective function (Z) based on the output of the at least one other neural sub-network (30a, ..., 30n) 2a ,...,Z 2n ), wherein the quality of object recognition performed by the at least one other neural sub-network (30a, ..., 30n) is determined by the at least one other objective function (Z 2a ,...,Z 2n ) to define, 3) Based on the evaluation of the first objective function (Z1) and based on the evaluation of the at least one other objective function (Z 2a ,...,Z 2n ) to determine the weights of the first neural sub-network (20), and 4) Based on the at least one other objective function (Z 2a ,...,Z 2n ) is evaluated to determine the weight of the at least one other neural sub-network (30a, ..., 30n).

2. The method according to claim 1, wherein Regions of the training data (3) with high entropy are less compressed than regions of the training data (3) with low entropy, wherein the training data (3) with low entropy includes image data with the same color and / or image data with recurring patterns.

3. The method according to any one of claims 1 to 2, wherein The other objective function (Z 2a ,...,Z 2n ) defines at least one of the following: identification of people, identification of vehicles, and identification of infrastructure.

4. A method for identifying an object by means of an object recognition system (100), comprising the steps of: - providing image data of the object (1); - compressing the image data (1) by means of a compression device (11a, ..., 11n) trained according to any one of the preceding claims 1 to 3 based on a first neural sub-network (20); and - performing an evaluation from the compressed image data (S) by means of an evaluation device (31a, ..., 31n) which has been trained according to any one of the preceding claims 1 to 3 based on at least one further neural sub-network (30a, ..., 30n) ka ,...,S kn ) for object recognition.

5. The method according to claim 4, wherein Simplified image data (1) with reduced entropy are generated from the image data (1), which are subsequently compressed by means of parameterized standard compression devices (11a, . . . , 11n).

6. The method according to claim 4, wherein: The image data (1) are compressed by means of parameterized standard compression means (11a, . . . , 11n).

7. A sensor device (10a, . . . , 10n) having a compression device (11a, . . . , 11n) configured by means of a method according to any one of claims 1 to 6.

8. An object recognition system (100) comprising a defined number of sensor devices (10a, ..., 10n) according to claim 7 which are functionally connected to one another, and a computing device (50) which is functionally connected to the sensor devices (10a, ..., 10n) via a data bus (4), the computing device (50) being configured to perform a computation from compressed image data (S k ) for object recognition.

9. The object recognition system according to claim 8, wherein: A first neural sub-network (20) is arranged on the sensor device (10a, . . . , 10n), and further neural sub-networks (30a, . . . , 30n) are arranged on the computing device (50). 10 . A computer program product comprising a computer program including instructions which, when executed by a computer, cause the computer to carry out the method according to claim 1 . 11 . A computer program product comprising a computer program including instructions which, when executed by a computer, cause the computer to carry out the method according to claim 4 . 12 . A machine-readable storage medium having stored thereon a computer program comprising instructions, which, when the computer program is executed by a computer, cause the computer to implement the method according to claim 1 .

Images