System and method for processing heterogeneous degraded images based on edge processor

Abstract

An edge processor-based heterogeneous degraded image processing system and method are disclosed. The present invention can restore images mixed with heterogeneous degradation in real time within a single processing structure based on an edge processor. It can maintain restoration quality and improve computational efficiency by determining the type and level of degradation of the input image according to signal characteristics and differentially applying quantization levels of restoration models for each degradation type. Furthermore, it can generate a final image through quality evaluation-based weighted synthesis after performing restoration by arranging multiple task-specific heads in parallel on a shared backbone network, and can operate platform-independently on various edge processors.

Description

Technology Field

[0001] The present invention relates to an edge processor-based heterogeneous degraded image processing system and method. More specifically, the invention relates to an edge processor-based heterogeneous degraded image processing system and method capable of maintaining restoration quality and improving computational efficiency by determining the type and level of degradation of an input image according to signal characteristics and differentially applying quantization levels of a restoration model for each degradation type to restore an image containing mixed heterogeneous degradation in real time within a single processing structure based on an edge processor, and capable of performing rapid degradation restoration through weighted synthesis using a shared backbone and a multi-head structure, and capable of operating platform-independently on various edge processors. Background Technology

[0002] Image acquisition devices such as CCTV surveillance systems, defense reconnaissance cameras, drone (UAV)-mounted cameras, and autonomous robots are operated in various degraded visual environments (DVE), such as night, adverse weather, and dusty environments.

[0003] Images acquired in such environments are characterized by the simultaneous or combined occurrence of multiple degradations, such as brightness reduction and color distortion due to low light, sensor noise, blur caused by camera or platform movement, and reduced visibility due to fog or haze.

[0004] Deep learning-based image restoration technology according to conventional technology uses individual models specialized for specific types of degradation, and when multiple degradations occur simultaneously in a real environment, individual models must be applied sequentially.

[0005] As a result, the amount of computation increases proportionally to the number of models, errors from each model propagate to subsequent stages leading to a deterioration in final restoration quality, and there is a problem in that the interaction between degradation types cannot be considered.

[0006] Recently studied All-in-One image restoration technology can process multiple degradations with a single network, but it has the problem that real-time inference is impossible in a lightweight NPU environment because it requires tens of millions of parameters.

[0007] In addition, conventional image restoration technology is designed based on GPU servers, so it fails to consider the memory and power constraints of edge devices and lacks consideration for NPU-compatible quantization, resulting in the problem that actual edge deployment is impossible.

[0008] In addition, image restoration technology according to the prior art consumes the same amount of computation regardless of the type or level of degradation of the input image, and thus performs unnecessary computations even when only specific degradation exists, resulting in the problem of wasting the limited resources of edge devices.

[0009] In addition, conventional image restoration technology is designed to be dependent on specific GPUs or frameworks, which presents a problem in that significant modifications to the model structure are required to port it to various NPU platforms.

[0010] Therefore, conventional technology fails to meet industrial requirements for providing identical performance across various hardware platforms, such as CCTVs, drones, and defense equipment, and has problems in that it fails to provide simultaneous processing of multiple degradations, real-time inference in edge device environments, adaptive allocation of computational resources, and compatibility across various platforms. Prior art literature

[0011] Korean Registered Patent Publication No. 10-2830406 (Title of Invention: Apparatus and Method for AI Decoding of Images, Registration Date: July 1, 2025)

[0012] 1. Li, B. et al., “All-in-One Image Restoration for Unknown Corruption”, IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.2. Potlapalli, V. et al., "PromptIR: Prompting for All-in-One Blind Image Restoration", Advances in Neural Information Processing Systems (NeurIPS), 2023.3. Cui, S. et al., "AdaIR: Adaptive All-in-One Image Restoration via Frequency Mining and Modulation", International Conference on Learning Representations (ICLR), 2025. The problem to be solved

[0013] To solve these problems, the present invention aims to provide an edge processor-based heterogeneous degraded image processing system and method capable of operating platform-independently on various edge processors, and capable of performing rapid degradation restoration through weighted synthesis via a shared backbone and multi-head structure, by determining the type and level of degradation of an input image based on signal characteristics to restore an image with mixed heterogeneous degradation in real time within an edge processor-based single processing structure, thereby maintaining restoration quality and improving computational efficiency, and enabling rapid degradation restoration. means of solving the problem

[0014] To achieve the above objective, one embodiment of the present invention is an edge processor-based heterogeneous degraded image processing system, comprising: a first processing method that generates multiple restored images by analyzing signal characteristics of an input image to calculate degradation weights for a plurality of degradation types and selecting different quantization levels for each degradation type according to the degradation weights; or a second processing method that generates multiple restored images by parallel inputting a common feature map extracted by processing the input image with a shared backbone network to a plurality of task-specific heads; and an image processing unit that generates a final restored image by weighted synthesizing the plurality of restored images by applying the degradation weights or the synthesis weights calculated by the second processing method.

[0015] In addition, the image processing unit according to the above embodiment is characterized by operating on an edge processor among an NPU (Neural Processing Unit), a GPU (Graphics Processing Unit), and an LPU (Language Processing Unit).

[0016] In addition, the image processing unit according to the above embodiment is characterized by generating a plurality of restored images by differentially selecting quantization levels to be applied to restoration models for each degradation type based on the degradation weight when the available computational resources in the edge processor are less than a preset reference value.

[0017] In addition, the image processing unit according to the above embodiment is characterized by generating multiple restored images by extracting a common feature map from the input image through a shared backbone network when the available computational resources in the edge processor are greater than or equal to the reference value, and inputting the common feature map into a plurality of task-specific heads to perform restoration processing in parallel.

[0018] In addition, the image processing unit according to the above embodiment is characterized by performing a combination of the first processing method and the second processing method, and generating a final restored image by weighted synthesis of the plurality of restored images by applying the degradation weight and the composite weight calculated by the second processing method.

[0019] In addition, the image processing unit according to the above embodiment is characterized by receiving the input image from an image acquisition unit comprising at least one of a CCTV camera, a drone-mounted camera, a defense surveillance camera, and an autonomous robot-mounted camera.

[0020] In addition, the image processing unit according to the above embodiment includes: a degradation determination unit that analyzes the signal characteristics of an input image and calculates degradation weights for each degradation type such as low light, noise, fog, and blur; an integrated processing unit that generates a plurality of images by selectively performing either a first processing method based on the degradation weights or a second processing method based on a shared backbone and a task-specific head, based on the available computational resources or real-time processing performance requirements of an edge processor; and an image synthesis unit that generates a final restored image by applying a synthesis weight calculated by the degradation weights or the second processing method to weight-synthesize the plurality of restored images. The image processing unit is characterized by operating on any one of an edge processor, such as an NPU, a GPU, and an LPU.

[0021] In addition, the degradation determination unit according to the above embodiment is characterized by comprising: a low-illumination determination unit that calculates a low-illumination degradation weight by analyzing the average brightness and frequency energy distribution of the input image; a noise determination unit that calculates a noise degradation weight by calculating the estimated peak signal-to-noise ratio of the input image; a fog determination unit that calculates a fog or haze degradation weight by analyzing the saturation information and frequency characteristics of the saturation of the input image; and a blur determination unit that calculates a blur degradation weight by analyzing the high-frequency region energy of the input image.

[0022] Additionally, the integrated processing unit according to the above embodiment is characterized by comprising: a control unit that, based on the available computational resources of the edge processor, selects a first processing method if the available computational resources are less than a preset threshold value, and selects a second processing method if the available computational resources are greater than or equal to the threshold value; a first processing unit that, when the first processing method is selected, generates a plurality of restored images by selecting a model with relatively high quantization precision for a restoration model of a relatively high degradation type or a model with relatively low quantization precision for a restoration model of a relatively low degradation type according to the degradation weight; and a second processing unit that, when the second processing method is selected, extracts a common feature map from an input image through the shared backbone network, generates a restored image by driving a plurality of task heads for each degradation type in parallel based on the common feature map, and calculates a composite weight by measuring the quality improvement of the generated restored image.

[0023] In addition, the first processing unit according to the above embodiment is characterized by comprising: a quantization level selection unit that, based on the degradation weight, selects a model with relatively high quantization precision for a restoration model of a degradation type with a relatively high degradation weight and selects a model with relatively low quantization precision for a restoration model of a degradation type with a relatively low degradation weight; a model storage unit that stores a plurality of restoration models corresponding to each of the plurality of degradation types and stores a plurality of quantization level models having different quantization precisions for each restoration model; and a restoration processing unit that generates a plurality of restored images by applying an input image to each restoration model of the selected quantization level.

[0024] In addition, the second processing unit according to the above embodiment is characterized by comprising: a shared backbone processing unit that extracts a common feature map from an input image through the shared backbone network; a parallel processing unit that receives the common feature map and generates a plurality of restored images by simultaneously processing task-specific heads corresponding to each of a plurality of degradation types in parallel; and a quality evaluation unit that calculates a composite weight by measuring the degree of improvement of the corresponding degradation type for each of the plurality of restored images.

[0025] In addition, the parallel processing unit according to the above embodiment is characterized by simultaneously performing low-light restoration, noise removal, fog removal, and blur removal processing in parallel through a plurality of independent threads within the edge processor.

[0026] In addition, the quality evaluation unit according to the above embodiment is characterized by measuring the low-illumination improvement based on the average brightness difference between the restored image and the input image, the noise improvement based on the difference in the estimated peak signal-to-noise ratio between the restored image and the input image, the fog improvement based on the degree of improvement in saturation and contrast of the restored image, and the blur improvement based on the high-frequency energy difference between the restored image and the input image.

[0027] In addition, one embodiment of the present invention is an edge processor-based heterogeneous degraded image processing method comprising: a) a step in which an image processing unit analyzes the signal characteristics of an input image and calculates a plurality of degradation weights for each degradation type; b) a step in which the image processing unit performs at least one of a first processing method in which a plurality of restored images are generated by selecting different quantization levels for each degradation type according to the degradation weights, or a second processing method in which a common feature map extracted by processing the input image with a shared backbone network is input in parallel to a plurality of task-specific heads to generate a plurality of restored images; and c) a step in which the image processing unit generates a final restored image by applying the degradation weights or the composite weights calculated by the second processing method to weight-synthesize the plurality of restored images.

[0028] In addition, the image processing method according to the above embodiment is characterized by operating on an edge processor among an NPU, a GPU, and an LPU.

[0029] Additionally, step a) according to the above embodiment is characterized by comprising: a step of calculating a low-illumination degradation weight by analyzing the average brightness and frequency energy distribution of the input image; a step of calculating a noise degradation weight by calculating the estimated peak signal-to-noise ratio of the input image; a step of calculating a fog or haze degradation weight by analyzing the saturation information and frequency characteristics of the saturation of the input image; and a step of calculating a blur degradation weight by analyzing the high-frequency region energy of the input image.

[0030] In addition, step b) according to the above embodiment is characterized by generating the plurality of restored images by differentially selecting a quantization level to be applied to a restoration model for each degradation type based on the degradation weight when the available computational resources in the edge processor are less than a preset threshold value.

[0031] In addition, step b) according to the above embodiment is characterized by generating the plurality of restored images by extracting a common feature map from the input image through a shared backbone network when the available computational resources in the edge processor are greater than or equal to the reference value, and inputting the common feature map into a plurality of task-specific heads to perform restoration processing in parallel.

[0032] In addition, step b) according to the above embodiment is characterized by performing the first processing method and the second processing method in combination, and generating a final restored image by weighted synthesis of the plurality of restored images by applying the degradation weight and the composite weight calculated by the second processing method.

[0033] In addition, the first processing method according to the above embodiment is characterized by comprising: a quantization level selection step of selecting a model with relatively high quantization precision for a restoration model of a degradation type with a relatively high degradation weight based on the degradation weight, and selecting a model with relatively low quantization precision for a restoration model of a degradation type with a relatively low degradation weight; and a step of generating a plurality of restored images by applying the input image to a restoration model pre-stored in a model storage unit corresponding to the selected quantization level.

[0034] In addition, the second processing method according to the above embodiment is characterized by comprising: a step of extracting a common feature map from the input image through a shared backbone network; a step of generating a plurality of restored images by inputting the common feature map into a task-specific head corresponding to each of a plurality of degradation types and processing them simultaneously in parallel; and a step of calculating the composite weight by measuring the degree of improvement of the corresponding degradation type for each of the plurality of restored images.

[0035] In addition, the step of simultaneous processing in parallel according to the above embodiment is characterized by simultaneously performing low-light restoration, noise removal, fog removal, and blur removal processing through a plurality of independent threads within the edge processor.

[0036] In addition, the step of calculating the composite weight according to the above embodiment is characterized by measuring and calculating the low-illumination improvement based on the average brightness difference between the restored image and the input image, the noise improvement based on the estimated peak signal-to-noise ratio difference, the fog improvement based on the degree of saturation and contrast enhancement, and the blur improvement based on the high-frequency energy difference.

[0037] In addition, one embodiment of the present invention may be configured as a computer-readable recording medium on which a program is recorded, comprising instructions for performing steps described in the above-described edge processor-based heterogeneous degraded image processing method when executed on a computer. Effects of the invention

[0038] The present invention has the advantage of being able to restore images with mixed heterogeneous degradation in real time in a single processing structure based on an edge processor, and to maintain restoration quality and improve computational efficiency by determining the type and level of degradation of the input image according to signal characteristics and differentially applying quantization levels of the restoration model for each degradation type.

[0039] In addition, the present invention has the advantage of reducing computational load by about 60% to 75% compared to conventional technology that sequentially applies a plurality of restoration models for each type of degradation, and improving processing speed by about 10 times or more.

[0040] In addition, the present invention has the advantage of minimizing the degradation of restoration quality while maintaining the number of parameters of the entire model at about 5M or less through adaptive quantization level selection based on degradation weights.

[0041] In addition, the present invention has the advantage of being able to generate a final image through quality evaluation-based weighted synthesis after performing restoration by arranging a plurality of task-specific heads in parallel on a shared backbone network, and to operate platform-independently on various edge processors.

[0042] In addition, the present invention can improve the performance of image analysis tasks such as object recognition and localization (SLAM) in the subsequent stage through image restoration, and has the advantage that the SLAM feature point extraction performance can be improved by about 190% compared to when low-light and fog images are fed into SLAM without restoration by performing low-light and fog restoration preprocessing. Brief explanation of the drawing

[0043] FIG. 1 is a block diagram illustrating an edge processor-based heterogeneous degraded image processing system according to an embodiment of the present invention. FIG. 2 is a block diagram illustrating the configuration of an image processing unit of an edge processor-based heterogeneous degraded image processing system according to an embodiment of FIG. 1. FIG. 3 is a block diagram illustrating the configuration of the degradation determination unit of the image processing unit according to the embodiment of FIG. 2. FIG. 4 is a block diagram illustrating the configuration of the integrated processing unit of the image processing unit according to the embodiment of FIG. 2. FIG. 5 is a block diagram illustrating the configuration of the first processing unit of the integrated processing unit according to the embodiment of FIG. 4. FIG. 6 is a block diagram illustrating the configuration of the second processing unit of the integrated processing unit according to the embodiment of FIG. 4. FIG. 7 is a flowchart illustrating an edge processor-based heterogeneous degraded image processing method according to an embodiment of the present invention. FIG. 8 is a flowchart illustrating the process of generating a restored image of an edge processor-based heterogeneous degraded image processing method according to an embodiment of FIG. 7. Specific details for implementing the invention

[0044] Hereinafter, the present invention will be described in detail with reference to preferred embodiments of the invention and the accompanying drawings, under the premise that identical reference numerals in the drawings refer to identical components.

[0045] Before describing specific details for the implementation of the present invention, it should be noted that configurations not directly related to the technical essence of the present invention have been omitted to the extent that they do not detract from the technical essence of the present invention.

[0046] Furthermore, terms or words used in this specification and claims should be interpreted in a meaning and concept consistent with the technical spirit of the invention, based on the principle that the inventor may define the concept of appropriate terms to best describe their invention.

[0047] In this specification, the expression that a part "includes" a certain component means that it does not exclude other components but may include additional components.

[0048] In addition, terms such as "...part," "...unit," and "...module" refer to a unit that processes at least one function or operation, and this can be classified as hardware, software, or a combination of both.

[0049] In addition, the term "at least one" is defined as a term including both singular and plural forms, and it is self-evident that even if the term "at least one" does not exist, each component may exist in the singular or plural form and may mean the singular or plural.

[0050] Hereinafter, a preferred embodiment of an edge processor-based heterogeneous degradation image processing system and method according to an embodiment of the present invention will be described in detail with reference to the attached drawings.

[0051] FIG. 1 is a block diagram illustrating an edge processor-based heterogeneous degraded image processing system according to an embodiment of the present invention; FIG. 2 is a block diagram illustrating the configuration of an image processing unit of an edge processor-based heterogeneous degraded image processing system according to an embodiment of FIG. 1; FIG. 3 is a block diagram illustrating the configuration of a degradation determination unit of an image processing unit according to an embodiment of FIG. 2; FIG. 4 is a block diagram illustrating the configuration of an integrated processing unit of an image processing unit according to an embodiment of FIG. 2; FIG. 5 is a block diagram illustrating the configuration of a first processing unit of an integrated processing unit according to an embodiment of FIG. 4; and FIG. 6 is a block diagram illustrating the configuration of a second processing unit of an integrated processing unit according to an embodiment of FIG. 4.

[0052] As shown in FIGS. 1 to 6, an edge processor-based heterogeneous degradation image processing system according to one embodiment of the present invention can restore an image mixed with heterogeneous degradation such as low light, noise, blur, and fog in real time at the edge processor.

[0053] In addition, the edge processor-based heterogeneous degraded image processing system simultaneously secures computational efficiency and quality by applying automatic identification based on signal characteristics and differential quantization by degradation type, and can generate a final image with improved restoration quality and computational efficiency through weighted synthesis using a shared backbone and a multi-head structure.

[0054] To this end, the edge processor-based heterogeneous degraded image processing system may be configured to include an image processing unit (100), an edge processor (200), and an image acquisition unit (300).

[0055] The image processing unit (100) can analyze the signal characteristics of the input image received from the image acquisition unit (300) and calculate a degradation weight for each of the multiple degradation types, such as low light, noise, fog, and blur.

[0056] Additionally, the image processing unit (100) can operate on any one of the edge processors (200) of the Neural Processing Unit (NPU), Graphics Processing Unit (GPU), and Language Processing Unit (LPU).

[0057] That is, the image processing unit (100) can perform deep learning-based processing tasks such as image restoration on the NPU, execute image restoration and deep learning models through parallel computation on the GPU, and can be configured to perform multimodal computation based on a common backbone network on the LPU.

[0058] Additionally, the image processing unit (100) can generate a plurality of restored images by performing at least one of a first processing method based on differential quantization based on degradation weights and a second processing method based on a shared backbone and task-specific head, taking into account the available computational resources or real-time processing requirements of the edge processor (200).

[0059] That is, if the available computational resources in the edge processor (200) are less than a preset reference value, the image processing unit (100) can generate multiple restored images by differentially selecting a quantization level to be applied to each restoration model for each degradation type based on a degradation weight.

[0060] Additionally, if the available computational resources in the edge processor (200) are greater than or equal to a preset reference value, the image processing unit (100) can generate multiple restored images through a second processing method in which a common feature map is extracted from an input image received from the image acquisition unit (300) through a shared backbone network, and the extracted common feature map is input into multiple task-specific heads to perform restoration processing in parallel.

[0061] Additionally, the image processing unit (100) can generate a final restored image by weightedly synthesizing a plurality of restored images generated by applying a degradation weight or a composite weight calculated by a second processing method.

[0062] Additionally, the image processing unit (100) may perform a combination of the first processing method and the second processing method, and may generate a final restored image by applying a degradation weight and a composite weight calculated by the second processing method to weight-synthesize multiple restored images.

[0063] To explain the image processing unit (100) in more detail, the image processing unit (100) may be configured to include a degradation detection unit (110), an integrated processing unit (120), and an image synthesis unit (130).

[0064] The degradation determination unit (110) can analyze the signal characteristics of the input image and calculate degradation weights for each type of degradation, such as low light, noise, fog, and blur.

[0065] That is, the degradation determination unit (110) can calculate a low-light degradation weight (w1), a noise degradation weight (w2), a fog degradation weight (w3), and a blur degradation weight (w4) for each type of degradation, and each weight can be output as a normalized value between 0 and 1.

[0066] To this end, the degradation detection unit (110) may be configured to include a low light detection unit (111), a noise detection unit (112), a fog detection unit (113), and a blur detection unit (114).

[0067] The low-illumination determination unit (111) can convert the input image into grayscale, calculate the average brightness, and additionally analyze the energy distribution in the frequency domain.

[0068] The low-illumination determination unit (111) determines that low-illumination degradation exists when the average brightness is less than a preset reference value (e.g., 50 / 255 or less) and the low-frequency energy ratio is greater than a certain value (e.g., 70%), and can calculate a low-illumination degradation weight (w1) in proportion to the low-illumination degradation level calculated based on the average brightness value and the low-frequency energy ratio.

[0069] The noise determination unit (112) calculates the estimated peak signal-to-noise ratio (ePSNR) for the input image, and, for example, uses a Laplacian variance-based noise estimation technique, and determines that noise degradation exists when the estimated peak signal-to-noise ratio is less than a preset reference value (ePSNR, for example, 25 dB), and can calculate a noise degradation weight (w2) inversely proportional to the estimated peak signal-to-noise ratio.

[0070] The fog detection unit (113) performs an average value of saturation and frequency analysis of saturation (e.g., discrete cosine transform, DCT) in the HSV color space of the input image, and if the average value of saturation is less than a preset reference value (e.g., 0.2) and the low-frequency energy ratio of saturation is greater than a certain value (e.g., 80%), it determines that fog or haze degradation exists and can calculate a fog degradation weight (w3) that is proportional to the average saturation and the low-frequency energy ratio.

[0071] The blur determination unit (114) applies energy in the high-frequency region to the input image, for example, a Laplacian filter or a Sobel filter, and then calculates the variance. If the calculated variance value is less than a preset reference value (e.g., 100), it determines that blur degradation exists, and can calculate a blur weight (w4) inversely proportional to the variance value.

[0072] Here, the low light degradation weight (w1), noise degradation weight (w2), fog degradation weight (w3), and blur degradation weight (w4) may be normalized and adjusted so that their sum is 1, or they may be maintained independently as values ​​between 0 and 1.

[0073] In addition, since the degradation detection process is performed using lightweight signal processing operations, only about 1% to less than 5% of the total computational load may be generated as overhead.

[0074] The integrated processing unit (120) can generate multiple images by selectively performing either a first processing method based on a differential quantization method based on degradation weights or a second processing method based on a shared backbone and a task-specific head, based on the available computational resources or real-time processing requirements of the edge processor (200).

[0075] To this end, the integrated processing unit (120) may be configured to include a control unit (121), a first processing unit (122), and a second processing unit (123).

[0076] The control unit (121) can monitor the available computational resources of the edge processor (200) (e.g., FLOPS, memory bandwidth, thread availability rate, etc.) in real time and compare them with preset reference values.

[0077] Here, available computing resources represent the spare computing power of the edge processor (200) in units of GFLOPS, and the reference value may be a parameter set during system design.

[0078] That is, if the available computational resources are less than the reference value, the control unit (121) determines that the available computational resources are insufficient and selects the first processing method, and if the available computational resources are greater than or equal to the reference value, it can select the second processing method.

[0079] For example, when a foggy road image at night is input from the image acquisition unit (300), the available computational resources of the edge processor (200) are measured to be less than a reference value, so the control unit (121) can select a first processing method.

[0080] Additionally, when a monitoring image of a backlight and ambient vibration environment is input from the image acquisition unit (300), the available computational resources of the edge processor (200) are measured to be greater than or equal to a reference value, so the control unit (121) can select a second processing method.

[0081] In addition, the control unit (121) can adjust the criteria for selecting the processing method even if there are sufficient resources, considering the real-time processing performance requirements, for example, the target frame rate, when processing of 30 frames or more is required, such as drone video.

[0082] The first processing unit (122) operates when the first processing method is selected by the control unit (121), and depending on the degradation weight, it can generate multiple restored images by selecting a model with relatively high quantization precision for a restoration model of a relatively high degradation type, or generate multiple restored images by selecting a model with relatively low quantization precision for a restoration model of a relatively low degradation type.

[0083] To this end, the first processing unit (122) may be configured to include a quantization level selection unit (122a), a model storage unit (122b), and a restoration processing unit (122c).

[0084] The quantization level selection unit (122a) can differentially determine the quantization level to be applied to the restoration model of each degradation type based on the degradation weight, such as INT4 (4-bit integer quantization), INT8 (8-bit integer quantization), and FP16 (16-bit floating-point).

[0085] INT4 can drastically reduce computational load and memory usage through a 4-bit integer quantization method, enabling the generation of fast restored images in ultra-low power environments, but the quality of the restored images may be somewhat lower due to the low representation precision.

[0086] INT8 is an 8-bit integer quantization method that can secure computational speed and memory efficiency without significantly losing accuracy when running a restoration model for each degradation type, making it suitable for real-time restoration in mobile devices or general GPU / NPU environments.

[0087] FP16 is a 16-bit floating-point method that can be used to generate restored images in a high-performance GPU environment, and can simultaneously secure accuracy and performance in high-resolution image restoration or multimodal operations.

[0088] Additionally, the quantization level selection unit (122a) can select a model with relatively high quantization precision for a restoration model of a relatively high degradation type based on the degradation weight transmitted from the degradation discrimination unit (110), and select a model with relatively low quantization precision for a restoration model of a degradation type with a relatively low degradation weight.

[0089] That is, the quantization level selection unit (122a) can select a high-precision model for a restoration model of a degradation type in which the degradation weight is a first reference value for selecting a restoration model, for example, 0.5 or higher.

[0090] Additionally, the quantization level selection unit (122a) can select a medium-precision model for a restoration model of a degradation type in which the degradation weight is greater than or equal to a second reference value for selecting a restoration model, for example, 0.2 and less than a first reference value.

[0091] Additionally, the quantization level selection unit (122a) may select a low-precision model for a restoration model of a degradation type in which the degradation weight is less than the second reference value, or omit the operation of the restoration model.

[0092] The quantization level selection unit (122a) enables efficient allocation of the computational resources of the limited edge processor (200) by performing high-quality restoration for the main degradation through such adaptive quantization level selection, while performing only minimal computation or omitting additional degradation.

[0093] The model storage unit (122b) is configured to store multiple restoration models corresponding to each of multiple degradation types and multiple quantization level models having different quantization precision for each restoration model, and can store pre-trained restoration models corresponding to low-light restoration, noise removal, fog removal, and blur removal.

[0094] Additionally, the model storage unit (122b) can store each restored model by converting it in advance to three quantization levels, INT4, INT8, and FP16, and classifying it into a low-precision model, a medium-precision model, and a high-precision model.

[0095] The low-precision model is a model with INT4 or low-precision INT8 quantization applied, which has somewhat lower restoration quality but the least amount of computation; the medium-precision model is a model with INT8 quantization applied, which can provide a balance between restoration quality and computation; and the high-precision model is a model with finely applied INT8 quantization or FP16 quantization applied, which has the highest restoration quality but the most computation.

[0096] The restoration processing unit (122c) can generate four restored images, such as a low-light restored image (R1), a noise-removed image (R2), a fog-removed image (R3), and a blur-removed image (R4), by applying the input image to each restoration model of the quantization level selected by the quantization level selection unit (122a) to perform low-light restoration, noise removal, fog removal, and blur removal.

[0097] Each reconstruction model can be composed of a hybrid architecture combining a Convolutional Neural Network (CNN) and a Transformer, and can be structured such that a CNN-based encoder extracts local features of the input image, a Transformer-based intermediate processing block learns long-range dependencies, and a CNN-based decoder generates the reconstructed image.

[0098] The second processing unit (123) operates when the second processing method is selected by the control unit (121), extracts a common feature map from the input image through a shared backbone network, and can generate a restored image by driving a plurality of task heads for each degradation type in parallel based on the common feature map.

[0099] In addition, the second processing unit (123) can calculate a composite weight by measuring the quality improvement of the generated restored image.

[0100] To this end, the second processing unit (123) may be configured to include a shared backbone processing unit (123a), a parallel processing unit (123b), and a quality evaluation unit (123c).

[0101] The shared backbone processing unit (123a) is configured to extract a Common Feature Map (CFM) from an input image through a shared backbone network, and can extract basic features that can be commonly utilized by multiple task heads.

[0102] Here, the shared backbone network has a hybrid structure in which CNN blocks and Transformer blocks are intersected, and can generate a common feature map by extracting low-level and intermediate-level features from the input image.

[0103] In addition, the shared backbone processing unit (123a) provides a common feature map so that multiple task-specific heads can utilize it simultaneously, thereby reducing redundant calculations and improving efficiency in the process of generating restored images corresponding to multiple degradation types.

[0104] The parallel processing unit (123b) receives a common feature map and can generate multiple restored images by having task heads corresponding to each of the multiple degradation types process them in parallel simultaneously.

[0105] That is, the parallel processing unit (123b) enables task-specific heads, such as a low light head (T1), a noise head (T2), a fog head (T3), and a blur head (T4), to be driven simultaneously in parallel through multiple independent threads within the edge processor (200).

[0106] Additionally, the parallel processing unit (123b) can simultaneously generate four restored images, such as a low-light restored image (R1'), a noise-removed image (R2'), a fog-removed image (R3'), and a blur-removed image (R4'), by executing each task head in parallel as multiple threads on the edge processor (200).

[0107] The quality evaluation unit (123c) can measure the improvement of the corresponding type of degradation by comparing four restored images with the input image, and can calculate the low light composite weight (w1'), noise composite weight (w2'), fog composite weight (w3'), and blur composite weight (w4') from this.

[0108] That is, the quality evaluation unit (123c) can measure the difference between the average brightness of the restored image and the average brightness of the input image, the low-light improvement based on the degree of histogram equalization, the noise improvement based on the difference in the estimated peak signal-to-noise ratio between the restored image and the input image, the fog improvement based on the degree of saturation improvement and contrast improvement of the restored image, and the blur improvement based on the difference in high-frequency energy between the restored image and the input image.

[0109] In addition, the quality evaluation unit (123c) may assign a high synthesis weight to a restored image with a high degree of improvement and a low synthesis weight to a restored image with a low degree of improvement.

[0110] The image synthesis unit (130) can generate a final restored image by weightedly synthesizing four restored images, such as low-light restored images (R1, R1'), noise removed images (R2, R2'), fog removed images (R3, R3'), and blur removed images (R4, R4').

[0111] When the first processing method is performed, the image synthesis unit (130) performs weighted synthesis using the degradation weight calculated by the degradation determination unit (110) as the synthesis weight, and the final restored image can be generated through w1*R1 + w2*R2 + w3*R3 + w4*R4.

[0112] Here, w1 is the low-light degradation weight, w2 is the noise degradation weight, w3 is the fog degradation weight, w4 is the blur degradation weight, R1 is the low-light restored image, R2 is the noise removed image, R3 is the fog removed image, and R4 is the blur removed image.

[0113] Additionally, the image synthesis unit (130) can perform weighted synthesis using the synthesis weights calculated by the quality evaluation unit (123c) when the second processing method is performed, and the final restored image can be generated through w1'*R1' + w2'*R2' + w3'*R3' + w4'*R4'.

[0114] Here, w1' is the low-light composite weight, w2' is the noise composite weight, w3' is the fog composite weight, w4' is the blur composite weight, R1' is the low-light reconstructed image, R2' is the noise removed image, R3' is the fog removed image, and R4' is the blur removed image.

[0115] In addition, if the sum of the weights is normalized to 1, it can be applied directly to the formula, and if it is not normalized, it can be converted into a weighted average and applied.

[0116] In addition, since weights are calculated by normalizing the quality improvement of each restored image, types of restored images with high actual restoration effects may be reflected more in the final result.

[0117] The edge processor (200) may be an edge AI chipset having a computational performance of about 8 TOPS to 26 TOPS, and may periodically measure available computational resources (free FLOPS, memory bandwidth, thread availability rate, etc.) internally and provide them to the image processing unit (100).

[0118] Additionally, the edge processor (200) may be composed of Hailo-8L (about 13 TOPS), Hailo-8 (about 26 TOPS), 'Qualcomm QRB5165', 'NVIDIA Jetson Orin NX', 'Ambarella CV72', etc., but is not limited thereto.

[0119] In addition, if the edge processor (200) is less than 8 TOPS (Tera Operations Per Second), the computational performance is insufficient, so the real-time processing speed in the composite degraded image restoration process is reduced, and the complex computation of the model cannot be handled, resulting in a degradation of the quality of the resulting image (residual noise, unremoved blur, etc.), and the parallel processing or weighted synthesis process of the restoration model is not performed smoothly, which may lower the overall system efficiency.

[0120] In addition, if the edge processor (200) is 26 TOPS or more, it may use excess resources compared to the actual required restoration work, which may cause problems such as increased power consumption and heat generation.

[0121] The image acquisition unit (300) is configured to acquire input images in real time and may be configured to include at least one of a CCTV camera, a drone-mounted camera, a defense surveillance camera, and an autonomous robot-mounted camera.

[0122] Additionally, the input video received through the video acquisition unit (300) can be transmitted to the video processing unit (100) in frame units, and may include RGB video of FHD (1920x1080) resolution, but is not limited thereto, and may include compressed or uncompressed video frames with an 8-bit depth.

[0123] The following describes a heterogeneous degraded image processing method based on an edge processor according to one embodiment of the present invention.

[0124] FIG. 7 is a flowchart illustrating an edge processor-based heterogeneous degraded image processing method according to an embodiment of the present invention, and FIG. 8 is a flowchart illustrating a process for generating a restored image of the edge processor-based heterogeneous degraded image processing method according to the embodiment of FIG. 7.

[0125] As shown in FIGS. 1 to 8, one embodiment of the present invention is an edge processor-based heterogeneous degraded image processing method, wherein an image processing unit (100) can receive an image signal (S100) from an image acquisition unit (300).

[0126] In addition, at step S100, the image processing unit (100) can analyze the signal characteristics of the input image and calculate degradation weights for each degradation type for low light, noise, fog, and blur.

[0127] In step S100, the image processing unit (100) can calculate a low-illumination degradation weight by analyzing the average brightness and frequency energy distribution of the input image, and can calculate a noise degradation weight by calculating the estimated peak signal-to-noise ratio (ePSNR) of the input image.

[0128] Additionally, in step S100, the image processing unit (100) can calculate a fog or haze degradation weight by analyzing the saturation information and frequency characteristics of the saturation of the input image, and can calculate a blur degradation weight by analyzing the high-frequency region energy of the input image.

[0129] After performing step S100, the image processing unit (100) can perform a comparative analysis with the available computational resources of the edge processor (200) and a preset reference value (S200).

[0130] Based on the analysis result of step S200, the image processing unit (100) can generate multiple restored images by performing at least one of a first processing method that generates multiple restored images by selecting different quantization levels for each degradation type according to degradation weights considering available computational resources or real-time processing requirements, or a second processing method that generates multiple restored images by processing input images with a shared backbone network and inputting common feature maps extracted to multiple task-specific heads in parallel (S300).

[0131] To explain step S300 in more detail, the image processing unit (100) can monitor the available computational resources of the edge processor (200) (e.g., FLOPS, memory bandwidth, thread availability rate, etc.) in real time and compare them with preset reference values ​​(S310).

[0132] In step S310, if the available computational resources in the edge processor (200) are less than a preset threshold value, the image processing unit (100) determines that the available computational resources are insufficient and selects a first processing method, and if the available computational resources are greater than or equal to the threshold value, it may select a second processing method.

[0133] If, as a result of the comparison in step S310, the available computational resources are less than a preset threshold value, the image processing unit (100) can select a first processing method and determine the type of degradation based on the degradation weight calculated in step S100 (S320).

[0134] Additionally, the image processing unit (100) can select a model with relatively high quantization precision for a restoration model of a relatively high degradation type according to the degradation weight, and select a model with relatively low quantization precision for a restoration model of a relatively low degradation type (S330).

[0135] In step S330, the image processing unit (100) can differentially determine the quantization level to be applied to the restoration model of each degradation type based on the degradation weight, such as INT4 (4-bit integer quantization), INT8 (8-bit integer quantization), and FP16 (16-bit floating-point).

[0136] Additionally, the image processing unit (100) can perform restoration processing to generate four restored images, such as a low-light restored image (R1), a noise-removed image (R2), a fog-removed image (R3), and a blur-removed image (R4), by applying the input image to each restoration model of the quantization level selected in step S330 (S340).

[0137] Additionally, if the available computational resources at the edge processor (200) in step S310 are greater than or equal to a preset reference value, the image processing unit (100) selects a second processing method and can extract a common feature map through a shared backbone network from an input image received from an image acquisition unit (300).

[0138] Additionally, the image processing unit (100) can input the extracted common feature map into a plurality of task-specific heads to perform restoration processing in parallel, and then generate a plurality of restored images.

[0139] To explain this in more detail, the image processing unit (100) can extract a common feature map from an input image received from an image acquisition unit (300) through a shared backbone network (S321).

[0140] In addition, at step S321, the image processing unit (100) provides a common feature map so that multiple task-specific heads can utilize it simultaneously, thereby reducing redundant calculations and improving efficiency in the process of generating restored images corresponding to multiple degradation types.

[0141] After performing step S321, the image processing unit (100) can receive a common feature map and generate multiple restored images by simultaneously processing task heads corresponding to each of the multiple degradation types in parallel (S331).

[0142] In step S331, the image processing unit (100) enables task-specific heads, such as a low-light head (T1), a noise head (T2), a fog head (T3), and a blur head (T4), to be driven simultaneously in parallel through multiple independent threads within the edge processor (200).

[0143] Additionally, in step S331, the image processing unit (100) can simultaneously generate four restored images, such as a low-light restored image (R1'), a noise-removed image (R2'), a fog-removed image (R3'), and a blur-removed image (R4'), by executing each task head in parallel as multiple threads on the edge processor (200).

[0144] After performing step S331, the image processing unit (100) compares four restored images with the input image to measure the improvement of the corresponding degradation type, and can calculate a low light composite weight (w1'), a noise composite weight (w2'), a fog composite weight (w3'), and a blur composite weight (w4') from this (S341).

[0145] In step S341, the image processing unit (100) can calculate a composite weight by measuring the low-light improvement based on the average brightness difference between the restored image and the input image, the noise improvement based on the estimated peak signal-to-noise ratio difference, the fog improvement based on the degree of saturation and contrast improvement, and the blur improvement based on the high-frequency energy difference.

[0146] Additionally, in step S341, the image processing unit (100) may assign a high synthesis weight to a restored image with a high degree of improvement and a low synthesis weight to a restored image with a low degree of improvement.

[0147] Additionally, the image processing unit (100) can generate a final restored image by weightedly synthesizing restored images such as low-light restored images (R1, R1'), noise removed images (R2, R2'), fog removed images (R3, R3'), and blur removed images (R4, R4') (S350).

[0148] In step S350, when the first processing method is performed, the image processing unit (100) can generate a final restored image by performing weighted synthesis using the degradation weight calculated by the degradation determination unit (110) as a synthesis weight.

[0149] In addition, when the second processing method is performed, the image processing unit (110) can generate a final restored image by performing weighted synthesis using the synthesis weights calculated in step S341.

[0150] Additionally, in step S300, the image processing unit (100) may perform a combination of the first processing method and the second processing method, and generate a final restored image by applying a degradation weight and a composite weight calculated by the second processing method to weight-synthesize multiple restored images.

[0151] In addition, one embodiment of the present invention may be configured as a computer-readable recording medium including an edge processor-based heterogeneous degraded image processing method, comprising a program having instructions for performing steps described in steps S100 to S300 when executed on a computer.

[0152] Therefore, images with mixed heterogeneous degradation can be restored in real time in a single processing structure based on an edge processor, and by determining the type and level of degradation of the input image according to signal characteristics and differentially applying quantization levels of the restoration model for each degradation type, the restoration quality can be maintained and computational efficiency improved.

[0153] In addition, restoration can be performed by arranging multiple task-specific heads in parallel on a shared backbone network, and a final image can be generated through quality evaluation-based weighted synthesis, and it can operate platform-independently on various edge processors.

[0154] As described above, although the present invention has been explained with reference to preferred embodiments, those skilled in the art will understand that various modifications and changes can be made to the present invention without departing from the spirit and scope of the invention as set forth in the following claims.

[0155] In addition, the drawing numbers described in the claims of the present invention are provided for clarity and convenience of explanation only and are not limited thereto, and in the process of describing the embodiments, the thickness of lines or the size of components shown in the drawings may be exaggerated for clarity and convenience of explanation.

[0156] Furthermore, the terms described above are defined in consideration of their functions in the present invention, and since these may vary depending on the intentions or practices of the user or operator, the interpretation of these terms should be based on the content throughout this specification.

[0157] Furthermore, it is obvious that a person skilled in the art to which the present invention pertains may make various modifications including the technical concept according to the present invention from the description of the present invention, even if not explicitly stated or described, and such modifications still fall within the scope of the rights of the present invention.

[0158] In addition, the above embodiments described with reference to the attached drawings are described for the purpose of explaining the present invention, and the scope of the present invention is not limited to these embodiments. Explanation of the symbols

[0159] 100: Image processing unit 110: Deterioration detection unit 111: Low-level discrimination unit 112: Noise discrimination unit 113: Fog detection unit 114: Blur detection unit 120 : Integrated processing unit 121 : Control unit 122: First processing unit 122a: Quantization level selection unit 122b : Model Storage Unit 122c : Restoration Processing Unit 123: Second Processing Unit 123a: Shared Backbone Processing Unit 123b: Parallel Processing Unit 123c: Quality Evaluation Unit 130: Image Synthesis Unit 200: Edge Processor 300 : Image acquisition unit

Claims

Claim 1 An edge processor-based heterogeneous degraded image processing system comprising an image processing unit (100) that selectively performs either a first processing method, which generates multiple restored images by analyzing the signal characteristics of an input image to calculate degradation weights for multiple degradation types and selecting different quantization levels to be applied to a restoration model for each degradation type according to the degradation weights based on the result of comparing available computational resources in an edge processor (200) with a preset reference value, or a second processing method, which generates multiple restored images by parallel inputting a common feature map extracted by processing the input image with a shared backbone network to multiple task-specific heads, and generates a final restored image by applying the degradation weights or the composite weights calculated by the second processing method to weight-synthesize the multiple restored images. Claim 2 An edge processor-based heterogeneous degraded image processing system, wherein, in claim 1, the image processing unit (100) generates a plurality of restored images by differentially selecting a quantization level to be applied to a restoration model for each degradation type based on the degradation weight when the available computational resources in the edge processor (200) are less than the reference value. Claim 3 An edge processor-based heterogeneous degraded image processing system according to claim 1, wherein the image processing unit (100) extracts a common feature map from the input image through a shared backbone network when the available computational resources in the edge processor (200) are greater than or equal to the reference value, and generates a plurality of restored images by inputting the common feature map to a plurality of task-specific heads and performing restoration processing in parallel. Claim 4 An edge processor-based heterogeneous degraded image processing system, wherein, in claim 1, the image processing unit (100) performs a combination of a first processing method and a second processing method, and applies a degradation weight and a composite weight calculated by the second processing method to weight-synthesize a plurality of restored images to generate a final restored image. Claim 5 In claim 1, the image processing unit (100) comprises: a degradation determination unit (110) that analyzes the signal characteristics of an input image and calculates degradation weights for each degradation type such as low light, noise, fog, and blur; an integration processing unit (120) that selectively performs either a first processing method based on the degradation weights or a second processing method based on a shared backbone and task-specific head to generate a plurality of images based on available computational resources or real-time processing requirements of an edge processor (200); and an image synthesis unit (130) that applies the degradation weights or the synthesis weights calculated by the second processing method to weight-synthesize the plurality of restored images to generate a final restored image; wherein the image processing unit (100) operates on an edge processor (200) that is one of an NPU (Neural Processing Unit), a GPU (Graphics Processing Unit), and an LPU (Language Processing Unit). Claim 6 In claim 5, the degradation determination unit (110) comprises: a low-illumination determination unit (111) that calculates a low-illumination degradation weight by analyzing the average brightness and frequency energy distribution of the input image; a noise determination unit (112) that calculates a noise degradation weight by calculating the estimated peak signal-to-noise ratio of the input image; a fog determination unit (113) that calculates a fog or haze degradation weight by analyzing the saturation information and frequency characteristics of the saturation of the input image; and a blur determination unit (114) that calculates a blur degradation weight by analyzing the high-frequency region energy of the input image; characterized in that it is an edge processor-based heterogeneous degradation image processing system. Claim 7 In claim 6, the integrated processing unit (120) comprises: a control unit (121) that selects a first processing method if the available computing resources of the edge processor (200) are less than a preset reference value, and selects a second processing method if the available computing resources are greater than or equal to the reference value; a first processing unit (122) that, when the first processing method is selected, generates a plurality of restored images by selecting a model with relatively high quantization precision for a restoration model of a relatively high degradation type or a model with relatively low quantization precision for a restoration model of a relatively low degradation type according to the degradation weight; and a second processing unit (123) that, when the second processing method is selected, extracts a common feature map from an input image through the shared backbone network, generates a restored image by driving a plurality of task heads for each degradation type in parallel based on the common feature map, and calculates a composite weight by measuring the quality improvement of the generated restored image. Claim 8 In claim 7, the first processing unit (122) comprises: a quantization level selection unit (122a) which, based on the degradation weight, selects a model with relatively high quantization precision for a degradation type restoration model with a relatively high degradation weight and a model with relatively low quantization precision for a degradation type restoration model with a relatively low degradation weight; a model storage unit (122b) which stores a plurality of restoration models corresponding to each of the plurality of degradation types and stores a plurality of quantization level models having different quantization precision for each restoration model; and a restoration processing unit (122c) which generates a plurality of restored images by applying an input image to each restoration model of the selected quantization level; an edge processor-based heterogeneous degradation image processing system. Claim 9 In claim 8, the second processing unit (123) comprises: a shared backbone processing unit (123a) that extracts a common feature map from an input image through the shared backbone network; a parallel processing unit (123b) that receives the common feature map and simultaneously processes a task-specific head corresponding to each of the plurality of degradation types in parallel to generate a plurality of restored images; and a quality evaluation unit (123c) that measures the degree of improvement of the corresponding degradation type for each of the plurality of restored images and calculates a composite weight; an edge processor-based heterogeneous degradation image processing system. Claim 10 a) a step in which an image processing unit (100) analyzes the signal characteristics of an input image and calculates a degradation weight for a plurality of degradation types; b) a step in which the image processing unit (100) selectively performs one of a first processing method to generate a plurality of restored images by selecting different quantization levels to be applied to a restoration model for each degradation type according to the degradation weight, based on the result of comparing available computational resources in an edge processor (200) with a preset reference value, or a second processing method to generate a plurality of restored images by parallel inputting a common feature map extracted by processing the input image with a shared backbone network to a plurality of task-specific heads; and c) a step in which the image processing unit (100) generates a final restored image by applying a synthesis weight calculated by the degradation weight or the second processing method to weight-synthesize the plurality of restored images. Claim 11 In claim 10, the above step a) comprises: a step of calculating a low-illumination degradation weight by analyzing the average brightness and frequency energy distribution of the input image; a step of calculating a noise degradation weight by calculating the estimated peak signal-to-noise ratio (ePSNR) of the input image; a step of calculating a fog or haze degradation weight by analyzing the saturation information and frequency characteristics of the saturation of the input image; and a step of calculating a blur degradation weight by analyzing the high-frequency region energy of the input image; characterized in that it comprises an edge processor-based heterogeneous degradation image processing method. Claim 12 In claim 10, the above step b) is characterized by generating the plurality of restored images by differentially selecting a quantization level to be applied to a restoration model for each degradation type based on the degradation weight when the available computational resources in the edge processor (200) are less than the reference value. Claim 13 In claim 10, the above step b) is characterized by generating a plurality of restored images by extracting a common feature map from the input image through a shared backbone network when the available computational resources in the edge processor (200) are greater than or equal to the reference value, and inputting the common feature map into a plurality of task-specific heads to perform restoration processing in parallel. Claim 14 An edge processor-based heterogeneous degraded image processing method, wherein in claim 10, step b) is performed by combining the first processing method and the second processing method, and a final restored image is generated by weighted synthesis of the plurality of restored images by applying the degradation weight and the composite weight calculated by the second processing method. Claim 15 In claim 10, the first processing method comprises: a quantization level selection step in which, based on the degradation weight, a model with relatively high quantization precision is selected for a restoration model of a degradation type with a relatively high degradation weight, and a model with relatively low quantization precision is selected for a restoration model of a degradation type with a relatively low degradation weight; and a step of generating a plurality of restoration images by applying the input image to a restoration model pre-stored in a model storage unit (122b) corresponding to the selected quantization level. Claim 16 In claim 10, the second processing method comprises: a step of extracting a common feature map from the input image through a shared backbone network; a step of generating a plurality of restored images by inputting the common feature map into task-specific heads corresponding to each of a plurality of degradation types and processing them simultaneously in parallel; and a step of calculating the composite weight by measuring the degree of improvement of the corresponding degradation type for each of the plurality of restored images; an edge processor-based heterogeneous degradation image processing method. Claim 17 A computer-readable recording medium having a program recorded thereon comprising instructions for performing steps described in the edge processor-based heterogeneous degraded image processing method of any one of claims 10 to 16 when executed on a computer.

Images