Auto-encoder for automotive vehicles with quantization

Abstract

A system (100) includes an encoder (110) and a decoder (120), wherein a first portion (111) of an auto-encoder convolutional neural network is integrated into the encoder and a second portion (122) of the auto-encoder convolutional neural network is integrated into the decoder. The encoder and the decoder are capable of communicating via a communication channel (130). A latent vector from the first portion of the neural network is quantized by the encoder for transmission over the communication channel, and then de-quantized by the decoder before processing by the second portion of the auto-encoder convolutional neural network. FIG. 1

Description

Title of the invention: Auto-encoder for motor vehicle with quantization

[0001] The present invention relates to the field of data compression, in particular image data. The invention applies in particular, but not exclusively, to data exchanged in or by a motor vehicle.

[0002] It is known to use auto-encoder systems to compress input data. An auto-encoder system is a convolutional artificial neural network obtained by unsupervised learning on a training data set, comprising an input layer and convolutional layers implemented in an encoder and inverted convolutional layers and an output layer implemented in a decoder, the input layer and the output layer having the same number of nodes, therefore the same number of dimensions.

[0003] The auto-encoder system thus comprises one or more hidden convolutional layers, each hidden layer having a number of dimensions less than the number of dimensions of the input layer and the output layer.

[0004] The hidden layer with the smallest number of dimensions, or central layer, is called the "code" or "latent vector" and the data represented in this central layer is thus a compressed version of the input data.

[0005] The autoencoder is trained by unsupervised learning so as to minimize the squared error between the input data and the output data from the output layer, given a number of dimensions of the central layer, or at a given compression level.

[0006] The central layer can thus be shared between the encoder and the decoder so as to exchange compressed data, thus making it possible to reduce the flow rate requirements and the quantity of data exchanged between the encoder and the decoder, while minimizing losses.

[0007] However, each dimension of the central layer is generally coded as a floating value over a large number of bits, in particular 32 bits. However, such coding is not compatible with many data links, in particular with a CAN bus which only accepts 8-bit integer data. For data links other than a CAN bus, it may also be required to compress the central layer data even more.

[0008] To this end, a first aspect of the invention relates to a method for encoding data, the method being implemented in an encoder storing layers of neurons comprising an input layer and at least one convolutional layer. hidden layer of an auto-encoder type convolutional neural network, the input layer comprising a first number of dimensions and said at least one hidden convolutional layer comprising a second number of dimensions, the second number of dimensions being less than the first number of dimensions, the encoding method comprising the following operations: - receiving input data of size equal to the first number of dimensions; - processing the input data by the layers of neurons so as to obtain a latent vector of size equal to the second number of dimensions, each dimension of the latent vector being coded according to a first coding format on N bits; - quantification of each dimension of the latent vector according to a second coding format on M bits, M being strictly less than N, to obtain a quantized latent vector; - transmission of the quantified latent vector over a communication channel compatible with the second data format.

[0009] Thus, the addition of a quantization makes it possible to transmit the quantized latent vector to any communication channel, including channels having restricted bit rates, or only accepting data on 8 bits, as is the case for example with a CAN bus, the quantized vector having a reduced size compared to the input data. Such an adaptation does not in practice induce any degradation, or a very slight degradation, of compression quality indicators of the input data.

[0010] A second aspect of the invention relates to a method for decoding compressed data, the method being implemented in a decoder storing layers of neurons comprising an output layer and at least one hidden layer of a convolutional neural network of the auto-encoder type, the output layer comprising a first number of dimensions and said at least one hidden convolutional layer comprising a second number of dimensions, the second number of dimensions being less than the first number of dimensions, the decoding method comprising the following operations: - reception by a communication channel of a quantized latent vector of size equal to the second number of dimensions, each dimension of the quantized latent vector being coded according to a second coding format on M bits, the communication channel being compatible with the second coding format; - de-quantization of the quantized latent vector into a latent vector estimate according to a first coding format, in which each dimension of the latent vector estimate is coded according to a first coding format on N bits, N being strictly greater than M; - processing of the latent vector by the layers of neurons so as to obtain output data of size equal to the first number of dimensions; - transmission of output data.

[0011] According to embodiments of the first aspect or the second aspect, the communication channel may be a CAN bus, the first N-bit coding format may be a 32-bit floating point value, and the second M-bit coding format may be an 8-bit integer value.

[0012] A CAN bus has the advantage of being a secure and inexpensive connection, also used in many devices, particularly in motor vehicles.

[0013] According to embodiments of the first aspect or the second aspect, the input data and the output data may be representative of automotive vehicle lighting photometry.

[0014] It is thus made possible to use the autoencoder convolutional neural network to transmit lighting commands, in the form of lighting photometry in particular, which is particularly advantageous in a motor vehicle.

[0015] According to embodiments of the first aspect or the second aspect, the method may further comprise a prior step of unsupervised training of the auto-encoder convolutional neural network on a training data set.

[0016] Such machine learning makes it possible to obtain high compression quality, with a low square error between the input data and the output data.

[0017] A third aspect of the invention relates to a computer program comprising instructions for implementing the method according to the first aspect or the second aspect of the invention, when these instructions are executed by a processor.

[0018] A fourth aspect of the invention relates to an encoder comprising: - a memory storing layers of neurons comprising an input layer and at least one hidden convolutional layer of an autoencoder convolutional neural network, said input layer comprising a first number of dimensions and said at least one hidden convolutional layer comprising a second number of dimensions, the second number of dimensions being less than the first number of dimensions; - a first interface capable of receiving input data of size equal to the first number of dimensions - a processor configured to process input data by the layers of neurons so as to obtain a latent vector of size equal to the second number of dimensions, each dimension of the latent vector being coded according to a first coding format on N bits, and quantize each dimension of the latent vector according to a second coding format on M bits, M being strictly less than N, to obtain a quantized latent vector; - a second interface capable of transmitting the quantified latent vector over a communication channel compatible with the second data format.

[0019] A fifth aspect of the invention relates to a decoder comprising: - a memory storing layers of neurons comprising an output layer and at least one hidden convolutional layer of an autoencoder type convolutional neural network, said output layer comprising a first number of dimensions and said at least one hidden convolutional layer comprising a second number of dimensions, the second number of dimensions being less than the first number of dimensions; - a first interface capable of receiving via a communication channel a quantized latent vector of size equal to the second number of dimensions, each dimension of the quantized latent vector being coded according to a second coding format on M bits, said communication channel being compatible with the second coding format; - a processor configured to de-quantize the quantized latent vector into a latent vector estimate according to a first coding format, in which each dimension of the latent vector estimate is coded according to a first coding format on N bits, N being strictly greater than M, and to process the latent vector by the layers of neurons so as to obtain output data of size equal to the first number of dimensions; - a second interface capable of transmitting output data.

[0020] A sixth aspect of the invention relates to a system comprising an encoder according to the fourth aspect of the invention and a decoder according to the fifth aspect of the invention.

[0021] According to one embodiment, the encoder can be integrated into a central control module of a motor vehicle and the decoder can be integrated into a lighting device of the motor vehicle.

[0022] Other characteristics and advantages of the invention will appear on examining the detailed description below, and the appended drawings in which:

[0023] [Fig-1] illustrates an auto-encoder system according to embodiments of the invention;

[0024] [Fig.2] is a diagram illustrating the steps of a data processing method input according to embodiments of the invention;

[0025] [Fig.3] is a diagram illustrating the steps of a data decoding method according to embodiments of the invention;

[0026] [Fig.4] illustrates a structure of an encoder according to embodiments of the invention;

[0027] [Fig.5] illustrates a structure of a decoder according to embodiments of the invention.

[0028] The description focuses on the characteristics that distinguish the methods, the system, the encoder and the decoder from those known in the state of the art.

[0029] [Fig.l] illustrates an auto-encoder system 100 according to embodiments of the invention.

[0030] The system 100 comprises an encoder 110 and a decoder 120, connected by a communication channel 130.

[0031] The encoder 110 may be integrated into equipment of a motor vehicle, such as a control module responsible for the lighting of the motor vehicle. Such a control module may be of the PCM type, for “Powertrain Control Module” in English, or ECU, for “Electronic Control Unit” in English, for example.

[0032] The decoder 120 can be integrated into equipment of a motor vehicle, such as a lighting device comprising lighting modules capable of performing lighting functions as a function of data communicated by the control module comprising the encoder 110. Preferably, at least one lighting module of the lighting device is a pixelated module, for example with a matrix of electroluminescent elements such as LEDs, with a matrix of micro-mirrors, of the DMD type for “Digital Micromirror Devices” in English, a monolithic source of electroluminescent elements on the same substrate, or any other technology allowing the production of a pixelated lighting beam. A monolithic source involves a plurality of electroluminescent semiconductor elements with submillimeter dimensions, epitaxially grown directly on a common substrate, the substrate generally being formed of silicon.Unlike conventional LED arrays, in which each elementary light source is an individually produced electronic component mounted on a substrate such as a printed circuit board (PCB), a monolithic source is to be considered as a single electronic component, during the production of which several areas of light-emitting semiconductor junctions are generated on a common substrate, in the form of an array.

[0033] The communication channel 130 can thus be a wired connection such as a CAN bus or an Ethernet connection. In the following, the example of a CAN bus is considered for illustrative purposes. It has the advantage of being a secure and inexpensive connection. However, a CAN bus requires data to be transmitted in the form of integers coded on 8 bits.

[0034] Alternatively, the encoder is integrated into a control module of the motor vehicle and the decoder 120 is integrated into a server remote from the motor vehicle. In this case, the communication channel comprises a wireless communication channel allowing the encoder to access an IP network in which the server is located. remote comprising the decoder 120. Such a wireless communication channel may be a cellular link of the 3G, 4G, 5G or any subsequent generation.

[0035] No restriction is attached to the communication channel 130, which can thus be a wired or wireless link. As will be better understood from reading what follows, most communication channels are constrained in terms of the flow rate and / or coding format of the data that they transmit.

[0036] According to the invention, the encoder and the decoder each comprise a part of a convolutional neural network of the auto-encoder type. The encoder 110 thus comprises a first part 112 of the auto-encoder convolutional neural network while the decoder 120 comprises a second part 122 of the auto-encoder convolutional neural network.

[0037] As previously explained, the autoencoder convolutional neural network is obtained by unsupervised learning on a training data set.

[0038] In the example considered here, in which the encoder is integrated into a PCM or ECU of a motor vehicle, and the decoder is integrated into a lighting device, the input data of the auto-encoder convolutional neural network are images representative of the light beam to be produced by the lighting device. Such images are also called photometries.

[0039] However, no restriction is attached to the input data which can be any type of image. For example, in the example in which the decoder 120 is located in a remote server, the input data can be images acquired by a camera of the vehicle.

[0040] An auto-encoder system is capable of compressing such input data. For this purpose, it is constructed by unsupervised learning on a training data set. It comprises an input layer 112 implemented in the encoder 110 and an output layer 124 implemented in a decoder 120, the input layer and the output layer having the same number of nodes, therefore the same number of dimensions.

[0041] The autoencoder system further comprises one or more hidden convolutional layers, each hidden convolutional layer having a number of dimensions less than the number of dimensions of the input layer 112 and the output layer 124.

[0042] The hidden convolutional layers having the smallest number of dimensions, or central layers, are able to exchange a "code" or "latent vector" and the data exchanged between these central layers is thus a compressed version of the input data.

[0043] The central layers can thus be shared between the encoder 112 and the decoder 123 so as to exchange compressed data, thus making it possible to reduce the throughput requirements and the quantity of data exchanged between the encoder and the decoder, while minimizing losses. The first part 111 thus comprises a central encoding layer 113 and the second part 122 comprises a central decoding layer 123. The central encoding layer 113 and the central decoding layer 123 are capable of exchanging a latent code or vector comprising a number of dimensions lower than the input data. The latent vector is thus a compressed version of the input data.

[0044] The autoencoder is trained by unsupervised learning so as to minimize the squared error between the input data and the output data from the output layer, given a number of dimensions of the central layers, or at a given compression level.

[0045] For this purpose, a training data set can be submitted to the auto-encoder. The training data set can comprise a set of images, such as vehicle lighting photometries in the example considered here. For each image of the training data set, the auto-encoder evaluates the quadratic error between the image submitted to the input layer and the image provided by the output layer, after processing by the auto-encoder, and varies the coefficients of its neurons as well as the number of neurons, or even the number of hidden convolutional layers, as a function of this quadratic error, while maintaining a constraint to obtain a latent vector having a given number of dimensions. The aim is thus to minimize the quadratic error by learning.

[0046] The latent vector is a compressed version of the input data, with a compression ratio CR according to the following formula: CR= (Nbits*hn_Size - NbitsLV*LVdim) / Nbits*hn_Size; where Nbits is the number of bits on which each pixel of the input image is encoded, Im_Size is the size of the image in number of pixels, NbitsLV is the number of bits encoding each dimension of the latent vector, which is fixed and generally equal to 32 bits, and LVdim is the number of dimensions of the latent vector.

[0047] More generally, Nbits*hn_Size represents the size in number of bits of the input data.

[0048] For a given image size, the compression rate CR can thus vary by varying the number of dimensions LVdim of the latent vector.

[0049] The following compression rates can in particular be obtained: - CR = 92% for LVdim = 516; - CR = 84% for LVdim = 1024; - CR = 52% for LVdim = 3072.

[0050] Thus, the lower the number of dimensions of the latent vector, the higher the rate of com CR pressure is high. The selection of a compression ratio may depend on quality indicators comparing the output data with the input data. Such indicators may include a peak signal-to-noise ratio, or PSNR, or a mean square error, or MSE, for example.

[0051] For example, a number of dimensions of the latent vector ensuring a PSNR greater than a given threshold, such as 30 for example, could be fixed.

[0052] The first part 111 and the second part 122 of the auto-encoder convolutional neural network are thus obtained, and can be implemented respectively in the encoder 110 and the decoder 120. The auto-encoder is thus produced by arranging a communication channel between the first part 111 and the second part 122.

[0053] Each dimension of the latent vector is generally a floating value coded on 32 bits, i.e. a first coding format. Such a value is incompatible with many communication channels, in particular with a CAN bus which can only transport 8-bit integer values.

[0054] The invention proposes, in order to overcome this problem, to integrate a quantization module 114 in the encoder 110 and a de-quantization module 121 in the decoder 120.

[0055] The quantization module 114 is capable of transforming each dimension of the latent vector from the first coding format on N bits, in particular on 32 bits in floating value, to a second coding format on M bits, M being less than N, with in particular M=8 to code an integer value, or alternatively M=16. The second coding format is a coding format compatible with the communication channel 130.

[0056] In this case, a second coding format corresponding to 8-bit integer values ​​is compatible with a CAN bus.

[0057] An example of quantization is given below in the specific embodiment corresponding to the passage from a 32-bit floating value to an 8-bit integer value, for illustrative purposes only.

[0058] Let a latent vector according to the following first coding format v= [0.1; 0.2; 0.3], with three dimensions, each dimension being a 32-bit floating value, between 0 (vmin) and 0.5 (vmax).

[0059] The quantization module 114 is therefore capable of providing a quantized latent vector, having LVdim dimensions, each dimension being coded according to the second coding format compatible with the communication channel 130.

[0060] The quantized latent vector is obtained as follows: Qv= rounded (v / qscale - qzero); where round() is the function to round a floating point value into an integer value; qscale = vmax - vmin; and qzero = 255 - vmax / qscale.

[0061] We thus obtain, in the example given above, the following quantified latent vector: Qv = [51; 102; 153],

[0062] The quantized latent vector Qv can thus be transmitted to the decoder 120 via the communication channel 130.

[0063] Upon receipt of the quantized latent vector Qv, the de-quantization module 121 applies the inverse transformation of the quantization module 112, in order to obtain an estimate of the latent vector, having LVdim dimensions and each dimension being according to the first coding format. The estimate of the latent vector is thus provided as input to the second part 122 of the auto-encoder convolutional neural network in order to determine output data. The convolutional neural network having been trained by machine learning, the output data are close to the input data, despite the difference between the latent vector from the first part 111 and the estimate of the latent vector from the de-quantization module 121.

[0064] The addition of a quantization module makes it possible to make the use of an auto-encoder compatible with most communication channels, in particular with a CAN bus. Moreover, such compatibility is permitted while maintaining high quality indicators for the compression of the input data.

[0065] The following compression levels can be obtained by the system of [Fig.l] with quantization: - CR = 98% for LVdim = 516; - CR = 96% for LVdim = 1024; - CR = 88% for LVdim = 3072.

[0066] In practice, compression quality indicators such as PSNR and MSE are maintained at levels as high, or nearly as high, as a non-quantizing autoencoder system.

[0067] Thus, the system according to the invention allows compatibility of an auto-encoder convolutional neural network with any type of communication channel, in particular with a CAN bus, by increasing the compression rate, while maintaining high compression quality indicators, which makes it possible to minimize losses on the output data.

[0068] [Fig.2] shows a method for processing input data, implemented by the encoder 110 described previously with reference to [Fig.l].

[0069] The method comprises a prior step 200 of storing the first part 111 of the auto-encoder convolutional neural network in a memory of the encoder 110. The preliminary step 200 may further comprise the learning method for obtaining the auto-encoder convolutional neural network, comprising the first part 111 and the second part 122.

[0070] The method further comprises a step 201 of receiving input data by the input layer 112. As described previously, the input data may be data representative of an image, such as photometry of a lighting device for a motor vehicle.

[0071] In a step 202, the input data are processed by the first part 111 and a latent vector is obtained at the output of the first central encoding layer 113, as previously described.

[0072] In a step 203, each dimension of the latent vector is quantized in the second coding format by the quantization module 114, as previously described, in order to obtain a quantized latent vector compatible with transmission on the communication channel 130.

[0073] In a step 204, the quantized latent vector is transmitted on the communication channel 130 by the encoder 110.

[0074] [Fig. 3] shows a data decoding method, implemented by the decoder 120 described previously with reference to [Fig. 1].

[0075] The method comprises a prior step 300 of storing the second part 122 of the auto-encoder convolutional neural network in a memory of the decoder 120. The prior step 200 may further comprise the learning method making it possible to obtain the auto-encoder convolutional neural network, comprising the first part 111 and the second part 122.

[0076] In a step 301, the de-quantization module 121 receives the quantized latent vector. In a step 302, the de-quantization module 121 determines an estimate of the latent vector from the quantized latent vector, each dimension of the estimate of the latent vector being in the first coding format, for example a 32-bit floating value. This step therefore makes it possible to switch from the second coding format to the first coding format.

[0077] At a step 303, the second part 122 of the autoencoder convolutional neural network processes the estimation of the latent vector in order to obtain output data, the output data having the same number of dimensions as the input data of step 201 of the method according to [Fig.2].

[0078] The output data may be transmitted by the decoder 120 at a step 304. For example, the decoder may transmit the output data to a memory for storage. Advantageously in the embodiment in which the encoder 110 is integrated in a PCM and the decoder 120 is integrated in a signaling device, the output data may be transmitted to a control module of light sources for producing photometry corresponding to the output data.

[0079] The invention also relates to a method comprising both the steps of the method of [Fig.2] but also the steps of the method of [Fig.3], the reception step 300 consisting of receiving the quantized latent vector transmitted in step 204.

[0080] [Fig.4] illustrates the structure of an encoder 110 according to embodiments of the invention.

[0081] The encoder 110 comprises a processor 401 configured to communicate unidirectionally or bidirectionally, via one or more buses or via a wired connection, with a memory 402 such as a memory of the “Random Access Memory” type, RAM, or a memory of the “Read Only Memory” type, ROM, or any other type of memory (Flash, EEPROM, etc.). Alternatively, the memory 402 comprises several memories of the aforementioned types. Preferably, the memory 402 is a non-volatile memory.

[0082] The memory 402 stores, permanently or temporarily, all of the data generated following the implementation of steps 201 to 204 of the data processing method described above, and can store the first part 111 of the convolutional neural network during step 200.

[0083] The processor 401 is capable of executing instructions, stored in the memory 402, for implementing steps 202 and 203 of the method illustrated with reference to [Fig.2]. Alternatively, the processor 401 can be replaced by a microcontroller designed and configured to carry out steps 202 and 203 of the method according to [Fig.3].

[0084] The quantization module 114 presented previously is thus implemented by the processor 401 or the microcontroller. As a further variant, a processor or a microcontroller dedicated to the quantization function is provided in addition to the processor 401, which thus performs the processing by the first part of the auto-encoder convolutional neural network.

[0085] The encoder 110 may comprise an input interface 403 capable of obtaining the input data during the step 201 described previously. No restriction is attached to the first input interface 403 which may be a wired interface for example, or alternatively wireless.

[0086] The encoder 110 may further comprise a second output interface 404 capable of transmitting the quantized latent vector, during the step 204 described previously. The second interface 404 is thus functionally connected to the communication channel 130.

[0087] [Fig.5] illustrates the structure of a decoder 120 according to embodiments of the invention.

[0088] The decoder 120 comprises a processor 501 configured to communicate unidirectionally or bidirectionally, via one or more buses or via a wired connection, with a memory 502 such as a memory of the “Random Access Memory” type, RAM, or a memory of the “Read Only Memory” type, ROM, or any other type of memory (Flash, EEPROM, etc.). Alternatively, the memory 502 comprises several memories of the aforementioned types. Preferably, the memory 502 is a non-volatile memory.

[0089] The memory 502 is capable of storing, permanently or temporarily, all of the data generated following the implementation of steps 301 to 304 of the data decoding method described above, and can store the second part 122 of the convolutional neural network during step 300.

[0090] The processor 501 is capable of executing instructions, stored in the memory 502, for implementing steps 302 and 303 of the method illustrated with reference to [Fig. 3]. Alternatively, the processor 501 can be replaced by a microcontroller designed and configured to carry out steps 302 and 303 of the method according to [Fig. 3].

[0091] The de-quantization module 121 presented previously is thus implemented by the processor 501 or the microcontroller. As a further variant, a processor or a microcontroller dedicated to the de-quantization function is provided in addition to the processor 501, which thus performs the processing by the second part 122 of the auto-encoder convolutional neural network.

[0092] The decoder 120 may comprise an input interface 503 capable of receiving the quantized latent vector, during the step 301 described previously. No restriction is attached to the first input interface 503, which is functionally connected to the communication channel 130 described previously.

[0093] The decoder 120 may further comprise a second output interface 504 capable of transmitting the output data, during the step 304 described previously.

[0094] The present invention is not limited to the embodiments described above as examples; it extends to other variants.

Claims

Claims

1. A method of encoding data, the method being implemented in an encoder (110) storing layers of a convolutional neural network (111) comprising an input layer (112) and at least one hidden convolutional layer (113) of a convolutional neural network of the auto-encoder type, said input layer comprising a first number of dimensions and said at least one hidden convolutional layer comprising a second number of dimensions, the second number of dimensions being less than the first number of dimensions, the encoding method comprising the following operations: - receiving (201) input data of size equal to the first number of dimensions; - processing (202) the input data by the layers of neurons so as to obtain a latent vector of size equal to the second number of dimensions, each dimension of the latent vector being coded according to a first coding format on N bits; - quantification (203) of each dimension of the latent vector according to a second coding format on M bits, M being strictly less than N, to obtain a quantized latent vector, the quantization (203) being carried out by a quantization module (114) in the coder (110); - transmission (204) of the quantized latent vector on a communication channel (130) compatible with the second data format.

2. A method of decoding compressed data, the method being implemented in a decoder (120) storing convolutional neural network layers (122) comprising, an output layer (124) and at least one hidden convolutional layer (123) of an autoencoder type convolutional neural network, said output layer comprising a first number of dimensions and said at least one hidden convolutional layer comprising a second number of dimensions, the second number of dimensions being less than the first number of dimensions, the decoding method comprising the following operations: - reception (301) by a communication channel (130) of a quantized latent vector of size equal to the second number of dimensions, each dimension of the quantized latent vector being coded according to a second coding format on M bits, said communication channel being compatible with the second coding format; - de-quantization (302) of the quantized latent vector into a latent vector estimate according to a first coding format, in which each dimension of the latent vector estimate is coded according to a first coding format on N bits, N being strictly greater than M, the quantization (302) being carried out by a dequantization module (121) in the decoder (120); - processing (303) of the latent vector by the layers of neurons so as to obtain output data of size equal to the first number of dimensions; - transmission (304) of the output data.

3. The method of claim 1 or 2, wherein the communication channel (130) is a CAN bus, wherein the first N-bit encoding format is a 32-bit floating point value and wherein the second M-bit encoding format is an 8-bit integer value.

4. Method according to one of the preceding claims, in which the input data and the output data are representative of a lighting photometry for a motor vehicle.

5. Method according to one of the preceding claims, further comprising a prior step (200; 300) of unsupervised learning of the auto-encoder convolutional neural network on a training data set.

6. Computer program comprising instructions for implementing the method according to any one of the preceding claims, when these instructions are executed by a processor (401; 501).

7. An encoder (110) comprising - a memory (402) storing convolutional neural network layers (111) comprising an input layer (112) and at least one hidden convolutional layer (113) of an auto-encoder convolutional neural network, said input layer comprising a first number of dimensions and said at least one hidden convolutional layer comprising a second number of dimensions, the second number of dimensions being less than the first number of dimensions; - a first interface (403) capable of receiving input data of a size equal to the first number of dimensions - a processor (401) configured to process input data by the layers of neurons so as to obtain a latent vector of size equal to the second number of dimensions, each dimension of the latent vector being coded according to a first coding format on N bits, and quantizing each dimension of the latent vector according to a second coding format on M bits, M being strictly less than N, to obtain a quantized latent vector; - a second interface (404) capable of transmitting the quantized latent vector on a communication channel (130) compatible with the second data format.

8. Decoder (120) comprising - a memory (502) storing convolutional neural network layers (122) comprising an output layer (123) and at least one hidden convolutional layer (124) of an autoencoder type convolutional neural network, said output layer comprising a first number of dimensions and said at least one hidden convolutional layer comprising a second number of dimensions, the second number of dimensions being less than the first number of dimensions; - a first interface (503) capable of receiving via a communication channel a quantized latent vector of size equal to the second number of dimensions, each dimension of the quantized latent vector being coded according to a second coding format on M bits, said communication channel being compatible with the second coding format;- a processor (501) configured to de-quantize the quantized latent vector into a latent vector estimate according to a first coding format, in which each dimension of the latent vector estimate is coded according to a first coding format on N bits, N being strictly greater than M, and to process the latent vector by the layers of neurons so as to obtain output data of size equal to the first number of dimensions; - a second interface (504) capable of transmitting output data.;

9. A system comprising an encoder (110) according to claim 7 and a decoder (120) according to claim 8.

10. System according to claim 9, wherein the encoder (110) is integrated into a central control module of a motor vehicle and wherein the decoder (120) is integrated into a lighting device of the motor vehicle. 16

Images