CN-121998881-A - Image reasoning method and device for construction and automatic driving of image quality conversion network

CN121998881ACN 121998881 ACN121998881 ACN 121998881ACN-121998881-A

Abstract

The embodiment of the invention relates to the technical field of image processing and automatic driving, and provides an image reasoning method and device for constructing an image quality conversion network and automatically driving, wherein the method comprises the steps of acquiring initial image data through a target image sensor; the method comprises the steps of obtaining initial image data, carrying out image data processing on the initial image data to obtain inferred image data and training image data, training an initial image quality conversion network through the inferred image data, verifying the trained initial image quality conversion network through the training image quality image data to obtain a verification result, and determining that the training of the initial image quality conversion network is completed when the verification result is passed to obtain a target image quality conversion network after the training is completed. Therefore, the expected performance of the automatic driving algorithm can be maintained, retraining of the perception network is avoided, and the cost for transferring the perception network to a new system is greatly reduced.

Inventors

PENG XIAOFENG
HUANG TENG
JING WEI
REN YIFAN
HE GUANGHUI

Assignees

北京辉羲智能科技有限公司

Dates

Publication Date: 20260508
Application Date: 20241107

Claims (10)

1. The construction method of the image quality conversion network is characterized by comprising the following steps: acquiring initial image data through a target image sensor, wherein the target sensor is a sensor arranged on a target vehicle integrating an automatic driving algorithm; Image data processing is carried out on the initial image data to obtain reasoning image quality image data and training image quality image data, the reasoning image quality image data is used for training a target image quality conversion network, and the training image quality image data is used for verifying the trained target image quality conversion network; training an initial image quality conversion network through the reasoning image quality image data, and verifying the trained initial image quality conversion network through the training image quality image data to obtain a verification result; and when the verification result passes, determining that the training of the initial image quality conversion network is completed, and obtaining a target image quality conversion network after the training is completed.
2. The method of claim 1, wherein said processing said initial image data to obtain inferred image quality image data and training image quality image data comprises: Pre-constructing a training data processing system for converting the image quality of the image data into a training image quality for verifying the image quality conversion network and an inference data processing system for converting the image quality of the image data into an inference image quality for training the image quality conversion network; and respectively inputting the initial image data into the training data processing system and the reasoning data processing system to process the image data, so as to obtain the reasoning image data for training the image quality conversion network and the training image data for verifying the image quality conversion network.
3. The method according to claim 2, wherein the training the initial image quality conversion network by the inferred image quality image data and verifying the trained initial image quality conversion network by the trained image quality image data, and obtaining the verification result, comprises: Inputting the reasoning image data to an initial image quality conversion network for deep learning to obtain a trained initial image quality conversion network; inputting the training image data to a trained initial image quality conversion network to obtain a network output result; And verifying the network output result and the training image quality image data to obtain a verification result.
4. A method according to claim 3, wherein said verifying said network output result and said training image quality image data to obtain a verification result comprises: and comparing the network output result with the image quality similarity of the training image data, and taking the image quality similarity as a verification result.
5. The method according to claim 4, wherein when the verification result passes, determining that the training of the initial image quality conversion network is completed, and obtaining a trained target image quality conversion network includes: And when the verification result is that the image quality similarity is greater than or equal to a preset similarity threshold, determining that the training of the initial image quality conversion network is completed, and taking the initial image quality conversion network after the training is completed as a target image quality conversion network.
6. The method of claim 5, wherein the acquiring initial image data by the target image sensor comprises: and acquiring initial image data with a preset number through a target image sensor, wherein the preset number is smaller than the average image number required by training a perception network in the automatic driving algorithm.
7. An image reasoning method of automatic driving, comprising: collecting image data to be inferred through a target image sensor arranged on a target vehicle; Inputting the image data to be inferred into the target image quality conversion network constructed according to any one of claims 1-6, so that the target image quality conversion network performs image quality conversion on the current image data to be inferred; and inputting the image data to be inferred after the image quality conversion into an automatic driving algorithm for image inference to obtain an image inference result.
8. An image quality conversion network constructing apparatus, comprising: the acquisition module is used for acquiring initial image data through the target image sensor; The processing module is used for processing the initial image data to obtain reasoning image quality image data and training image quality image data, wherein the reasoning image quality image data is used for training a target image quality conversion network, and the training image quality image data is used for verifying the trained target image quality conversion network; the network training module is used for training an initial image quality conversion network through the reasoning image quality image data and verifying the trained initial image quality conversion network through the training image quality image data to obtain a verification result; And the determining module is used for determining that the training of the initial image quality conversion network is completed when the verification result passes, and obtaining a target image quality conversion network after the training is completed.
9. An image reasoning apparatus for automatic driving, comprising: the acquisition module is used for acquiring image data to be inferred through a target image sensor arranged on a target vehicle; The processing module is used for inputting the image data to be inferred into a target image quality conversion network so that the target image quality conversion network can convert the image quality of the current image data to be inferred; The reasoning module is used for inputting the image data to be reasoning after the image quality conversion into an automatic driving algorithm for image reasoning to obtain an image reasoning result.
10. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that the computer program, when executed by a processor, implements the construction method of the image quality conversion network according to any one of claims 1 to 6 and the image reasoning method of automatic driving according to claim 7.

Description

Image reasoning method and device for construction and automatic driving of image quality conversion network Technical Field The invention relates to the technical field of image processing and automatic driving, in particular to an image reasoning method and device for constructing an image quality conversion network and automatically driving. Background The automatic driving technology is an important technology in the field of intelligent automobiles at present, and is also the direction of intensive research of various manufacturers at present. After the automobile enters an automatic driving mode, surrounding environment or road condition information needs to be collected and analyzed in real time, and an automatic driving algorithm is needed. The automatic driving algorithm uses a sensing network to calculate surrounding environment or road condition images acquired in real time. If the input image data to be inferred is different from the training image data applied during algorithm training, the performance of the perceived network is degraded. When the automatic driving algorithm is transferred from one system to another, for example, products are shifted from research and development to the ground, because of the change of cameras and the difference of data acquisition systems, the image data for early training and the image data of reasoning input shot by a new system are difficult to be consistent in image quality, and thus, the phenomenon of reduced perceived network performance inevitably occurs. In order to solve the above problems, the existing method is to use a new system to re-collect a large amount of image data as a training set and re-train the sensing network in the new system. The existing method needs more than tens of millions of images to be re-acquired, and the training time is long, so that a great deal of manpower, material resources and time are consumed. Disclosure of Invention The invention provides a method and a device for constructing an image quality conversion network and image reasoning of automatic driving, which are used for solving the defect that the automatic driving algorithm needs to be retrained time and labor consumption in the prior art after being migrated, realizing that the expected performance of the automatic driving algorithm can be kept, avoiding the retrained of a perception network and greatly reducing the cost of migrating the perception network to a new system. The invention provides a construction method of an image quality conversion network, which comprises the following steps: acquiring initial image data through a target image sensor, wherein the target sensor is a sensor arranged on a target vehicle integrating an automatic driving algorithm; Image data processing is carried out on the initial image data to obtain reasoning image quality image data and training image quality image data, the reasoning image quality image data is used for training a target image quality conversion network, and the training image quality image data is used for verifying the trained target image quality conversion network; training an initial image quality conversion network through the reasoning image quality image data, and verifying the trained initial image quality conversion network through the training image quality image data to obtain a verification result; and when the verification result passes, determining that the training of the initial image quality conversion network is completed, and obtaining a target image quality conversion network after the training is completed. According to the construction method of the image quality conversion network provided by the invention, the method further comprises the following steps: Pre-constructing a training data processing system for converting the image quality of the image data into a training image quality for verifying the image quality conversion network and an inference data processing system for converting the image quality of the image data into an inference image quality for training the image quality conversion network; and respectively inputting the initial image data into the training data processing system and the reasoning data processing system to process the image data, so as to obtain the reasoning image data for training the image quality conversion network and the training image data for verifying the image quality conversion network. According to the construction method of the image quality conversion network provided by the invention, the method further comprises the following steps: Inputting the reasoning image data to an initial image quality conversion network for deep learning to obtain a trained initial image quality conversion network; inputting the training image data to a trained initial image quality conversion network to obtain a network output result; And verifying the network output result and the training image quality image data to obtain a verification result. According to