Multi-Batch super-resolution image frame selection method and system based on multi-core processing unit

Abstract

The invention provides a multi-Batch super-resolution image frame selection method and system based on a multi-core processing unit, and the method comprises the following steps: carrying out the super-resolution reasoning of a high-resolution image of the same frame, and outputting a corresponding super-resolution reasoning result through each super-resolution processing unit; performing super-resolution quality evaluation on each super-resolution reasoning result based on a super-resolution quality evaluation unit, and obtaining a corresponding super-resolution quality evaluation result; and outputting a final super-resolution result based on a super-resolution quality evaluation result to complete image frame selection. According to the method, the parallel computing capability of the multi-core super-resolution processing unit is fully exerted through the multi-Batch processing strategy, the processing efficiency of the super-resolution model is greatly improved, the robustness and the real-time performance of the super-resolution processing quality are ensured, and meanwhile accurate quality evaluation of the super-resolution result is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and in particular to a multi-Batch super-resolution image frame selection method and system based on a multi-core processing unit. BACKGROUND

[0002] Existing super-resolution algorithms (i.e., "super-resolution algorithms") have been widely applied in fields that require image quality improvement and detail reconstruction, such as video surveillance, intelligent security, remote sensing image processing, medical image enhancement, etc. They mainly rely on deep learning models for high-precision inference calculation, and extract features from local image information to reconstruct higher-quality images.

[0003] At the same time, many super-resolution algorithms have been optimized and ported to embedded platforms and AI computing platforms to improve inference efficiency and real-time processing capability. However, in scenarios using embedded devices or low-power AI computing platforms, existing super-resolution algorithms have the following limitations in actual application:

[0004] 1. Inference load cannot be evenly distributed to all available computing cores, resulting in low utilization of computing resources and waste of computing power.

[0005] 2. Existing super-resolution algorithms usually use whole-image input for one-time inference when processing large-size images, which occupies a large amount of video memory resources, seriously affecting system stability and task continuity, and the inference execution time is too long.

[0006] 3. The quality of the results of existing super-resolution algorithms is unstable and lacks optimization mechanisms.

[0007] 4. Existing super-resolution image quality evaluation methods mainly rely on traditional full-reference evaluation indicators such as PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index), but these indicators have obvious limitations in evaluating image detail reconstruction quality and realistic perception effect, and cannot provide a basis for performance optimization of super-resolution algorithms. SUMMARY

[0008] The present application relates to the technical field of image processing, and in particular to a multi-Batch super-resolution image frame selection method and system based on a multi-core processing unit.

[0009] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0010] In one aspect, a multi-Batch super-resolution image frame selection method based on a multi-core processing unit is provided, which comprises the following steps:

[0011] Super-resolution inference is performed on the same frame of high-resolution images, and each super-processing unit outputs a corresponding super-resolution inference result;

[0012] Each super-resolution inference result is evaluated based on a super-resolution quality evaluation unit, and a corresponding super-resolution quality evaluation result is obtained;

[0013] The final super-resolution result is output based on the super-resolution quality evaluation result to complete image frame selection.

[0014] Preferably, the high-resolution image is obtained by super-resolution processing of a single low-resolution image, which comprises the following steps:

[0015] The single low-resolution image is divided into n image blocks;

[0016] n Batch is set, and each Batch performs super-resolution inference on one image block and outputs the super-resolution processing result of each image block to complete the multi-Batch inference task;

[0017] The super-resolution processing results of all image blocks are spliced to obtain a complete high-resolution image.

[0018] Preferably, the super-resolution inference of all Batch on the image blocks is performed in parallel.

[0019] Preferably, n Batch is set on the same multi-Batch inference unit, and each Batch performs super-resolution inference on the image blocks by using a corresponding super-processing model.

[0020] Preferably, N independent super-processing units are integrated in the same platform.

[0021] Preferably, the super-resolution inference of all super-processing units on the same frame of high-resolution images is performed in parallel.

[0022] Preferably, at least two super-processing units use different super-processing models for super-resolution inference.

[0023] Preferably, the final super-resolution result is output based on the super-resolution quality evaluation result to complete image frame selection, which specifically comprises the following steps:

[0024] All super-resolution quality evaluation results are converted into binary Mask masks;

[0025] The super-resolution inference result corresponding to the super-resolution quality evaluation result with Mask mask of 1 is taken as the final super-resolution result and is output.

[0026] Preferably, the final super-resolution result is output based on the super-resolution quality evaluation result to complete the image frame selection, which specifically includes the following steps:

[0027] Convert all super-resolution quality assessment results into binary masks;

[0028] The super-resolution inference result corresponding to the super-resolution quality assessment result with Mask = 1 is used as the candidate super-resolution result;

[0029] If no new super-resolution quality assessment result is generated within the predetermined time period, there is no need to update the candidate super-resolution result. The current candidate super-resolution result is directly used as the final super-resolution result and output;

[0030] If a new super-resolution quality assessment result with a mask of 1 is generated within a predetermined time period, the super-resolution inference result corresponding to the super-resolution quality assessment result is used as the current super-resolution result to replace the original super-resolution result to be selected, and the super-resolution result to be selected after the replacement is completed is output.

[0031] On the other hand, a multi-batch super-resolution image frame selection system is also provided, which includes:

[0032] N super-resolution processing units, which are used to perform super-resolution inference on the same frame of high-resolution image, and each super-resolution processing unit outputs a corresponding super-resolution inference result;

[0033] a super-resolution quality evaluation unit, which performs super-resolution quality evaluation on each super-resolution inference result based on the super-resolution quality evaluation unit, and obtains a super-resolution quality evaluation result corresponding to each super-resolution inference result;

[0034] And, a super-resolution result output unit outputs the final super-resolution result based on the super-resolution quality evaluation result to complete the image frame selection.

[0035] In summary, the present invention has the following beneficial effects compared with the prior art:

[0036] The present invention integrates the super-resolution processing unit and the super-resolution quality assessment unit on the same platform. Through the multi-batch processing strategy, the parallel computing capability of the multi-core super-resolution processing unit is fully utilized, which greatly improves the processing efficiency of the super-resolution model, ensures that the image processing process can also run efficiently in a resource-constrained environment, and ensures the robustness and real-time performance of the super-resolution processing quality.

[0037] Meanwhile, the application carries out super-resolution quality evaluation by deploying a no-reference evaluation model such as MANIQA, and combines mask marking and outputs the super-resolution image with the best super-resolution processing effect, so as to realize accurate quality evaluation of the super-resolution result, and update the selected super-resolution result, thereby adaptively coping with the change of image quality and improving the processing effect and stability of the super-resolution model. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 A step flowchart of the multi-Batch super-resolution image selection method based on the multi-core processing unit in the application;

[0039] Figure 2 A flowchart of step S1 in the application;

[0040] Figure 3 A structural diagram of the super-resolution processing unit and the super-resolution quality evaluation unit in the application;

[0041] Figure 4 Super-resolution effects of the benchmark model, the benchmark_task sync model, and the m8c32_paired_div2k_relu model on 36 images;

[0042] Figure 5 Evaluation results of the super-resolution quality evaluation of the super-resolution inference result by PSNR and SSIM;

[0043] Figure 6 A flowchart of step S4 in the application;

[0044] Figure 7 A structural diagram of the multi-Batch super-resolution image selection system in the application. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0046] Embodiment 1

[0047] Figure 1 As shown, the embodiment provides a multi-Batch super-resolution image selection method based on a multi-core processing unit, which includes the following steps:

[0048] S1, super-resolution processing is performed on a single-frame low-resolution image to output a high-resolution image, which specifically includes the following steps:

[0049] The single-frame low-resolution image is divided into n image blocks (i.e. Figure 2 1-1, 1-2,..., 1-n) in the embodiment, the size of each image block is the same;

[0050] n batches are set, and each batch corresponds to super-resolution inference on one image block, and the super-resolution processing result of each image block is output to complete a multi-batch inference task; specifically, in the embodiment, the n batches are set on the same multi-batch inference unit, each batch performs super-resolution inference on the image block through the corresponding super-resolution processing model, and the super-resolution inference of all batches on the image block is performed in parallel, so that the multi-batch inference task can be completed by only one multi-batch inference unit, and the multi-batch inference unit includes an NPU (Neural network Processing Unit, neural network processing unit), in the embodiment, n is a positive integer greater than or equal to 2;

[0051] The super-resolution processing results of all image blocks are spliced to obtain a complete high-resolution image, and the high-resolution image is output, thereby realizing super-resolution processing on the low-resolution image;

[0052] Therefore, in this step, the low-resolution image is divided into multiple image blocks, and the multiple image blocks are distributed to different batches for parallel inference, so that all computing cores of the single-core processing unit (i.e. multiple batches set on the same multi-batch inference unit) are fully activated through multi-batch inference, the computing potential of the single-core inference unit is fully utilized, the parallel computing capability of the hardware is utilized to avoid the calculation bottleneck of single batch when processing large images, the utilization rate of hardware resources is greatly improved while avoiding waste of computing resources, and the inference speed is accelerated, and the overall image processing efficiency is improved;

[0053] S2, super-resolution inference is performed on the same frame of high-resolution image by N independent super-resolution processing units, and each super-resolution processing unit outputs a corresponding super-resolution inference result, wherein the super-resolution inference result is a super-resolution image obtained by the super-resolution processing unit performing super-resolution inference on the high-resolution image;

[0054] Specifically, in the embodiment, N independent super-resolution processing units (i.e. Figure 3SR Model1, SRModel2,..., SR Modex-1) in the same platform (such as a chip), and each super-resolution processing unit can be an NPU, and all super-resolution processing units perform super-resolution inference on the same frame of high-resolution image in parallel, and N is a positive integer greater than or equal to 2;

[0055] Meanwhile, each super-resolution processing unit has the ability to independently perform deep learning inference tasks, and each super-resolution processing unit uses a corresponding super-resolution processing model to perform super-resolution inference on the same frame of high-resolution image, and at least two super-resolution processing units use different super-resolution processing models when performing super-resolution inference, and preferably, each super-resolution processing unit uses a different super-resolution processing model when performing super-resolution inference.

[0056] Among them, the super-resolution processing model is different, including one or more of different training strategies of the super-resolution processing model, different parameter configurations of the super-resolution processing model, and different training data sets of the super-resolution processing model.

[0057] In the prior art, when processing large-size images through a super-resolution algorithm, the entire image is usually directly input into a super-resolution model for one-time inference. However, due to the extremely high pixel density of large-size images, a large amount of video memory resources will be occupied, especially under the memory limited condition of embedded devices or low-power AI platforms, which can easily cause video memory overflow, buffer jam and other problems, seriously affecting system stability and task continuity. At the same time, the calculation overhead of whole-image inference is high, and the inference process is often accompanied by large-scale convolution operation and feature mapping, which greatly prolongs the inference execution time, making it difficult to meet the real-time processing requirements of high frame rate and low delay. Especially in the continuous input scene such as ultra-high definition video (such as 4K / 8K content) or high-speed image stream, the above problems are particularly prominent. In addition, whole-image inference lacks flexibility, which limits the algorithm's ability in resource dynamic scheduling.

[0058] In this step, each independent super-resolution processing unit can independently run a corresponding super-resolution processing model to perform N times of super-resolution inference on the same frame of high-resolution image in parallel, thereby fully utilizing the parallel computing advantage of the multi-core processing unit, significantly reducing the memory requirement, speeding up the inference, and avoiding memory overflow and calculation bottleneck problems.

[0059] At the same time, the super-resolution processing model used by the super-resolution processing unit when performing super-resolution inference is not a repeated copy or reuse of the same model, but can be partially different or completely different, and ultimately N super-resolution inference results that are partially the same or completely different can be obtained, providing a variety of candidate results for subsequent image quality screening and optimization, and further improving the super-resolution effect and output stability of the system.

[0060] S3, based on the super-resolution quality evaluation unit (i.e.Figure 3 The IQA Model in the above example performs a super-resolution quality assessment on each super-resolution inference result and obtains the super-resolution quality assessment result corresponding to each super-resolution inference result (i.e. Figure 3 Result-1, Result-2,..., Result-x-2, Result-x-1) in

[0061] Similarly, the super-resolution quality assessment unit may also be an NPU, which completes the super-resolution quality assessment of each super-resolution inference result based on the image quality assessment model, and all super-resolution processing units and super-resolution quality assessment units may be integrated into the same platform (such as a chip). Furthermore, the image quality assessment model in this embodiment includes a no-reference assessment model, such as one or more of MANIQA, CLIP-IQA, and MUSIQ.

[0062] For example, in this embodiment, the super-resolution quality assessment unit can complete the super-resolution quality assessment of each super-resolution inference result based on MANIQA, and further, Figure 4 As shown, in this embodiment, three super-resolution models (i.e., the benchmark model, the benchmark_tasksync model, and the m8c32_paired_div2k_relu model) are selected to perform super-resolution processing on 36 images (i.e., "images"). Through visual observation and other equipment and evaluation programs, it is pre-determined that the benchmark model has the worst super-resolution effect, the benchmark_tasksync model has a medium super-resolution effect, and the m8c32_paired_div2k_relu model has the best super-resolution effect. The images after super-resolution processing by the three models are scored by the MANIQA model. According to the results, the super-resolution results of the benchmark model, the benchmark_tasksync model, and the m8c32_paired_div2k_relu model are 9.2446, 9.9072, and 10.7794, respectively. The results are consistent with the above-determined super-resolution effect conclusions. This effectively proves that the MANIQA scoring system can evaluate the quality of super-resolution results.

[0063] At the same time, the full reference evaluation model in the existing technology, namely PSNR and SSIM, is used to evaluate the super-resolution quality of each super-resolution inference result. The evaluation results are as follows: Figure 5As shown in the table, Sharpness1, Noise1, Detail1 are the super-resolution quality evaluation results obtained after step S3 is executed when only 1 batch is set in step S1, that is, sharpness, noise, and detail, Sharpness2, Noise2, Detail2 are the super-resolution quality evaluation results obtained after step S3 is executed when 2 batches are set in step S1, and "picture name" is the super-resolution inference result in step S3. The super-resolution inference result is evaluated by using PSNR and SSIM.

[0064] As can be seen from Figure 5 It can be seen that the super-resolution effect of setting multiple batches is better than that of single batch. At the same time, PSNR and SSIM have obvious limitations in evaluating image detail reconstruction quality and real perception effect. The reason is that PSNR is essentially a pixel error-based measurement method, which pays more attention to the average error in numerical value and ignores the perception difference of human eyes to local details, texture distribution and visual sensitive areas of images. Although SSIM introduces structural information contrast to a certain extent, it still mainly focuses on low-frequency information of images and cannot effectively reflect the recovery quality of high-frequency texture, edge sharpness and microstructure. Especially when facing images with complex texture, rich details or existing degradation blur, the scores of PSNR and SSIM may appear "high quality score and low quality score" or "low quality score and high quality score" evaluation deviation, which further affects the algorithm optimization direction and model iteration strategy.

[0065] In this embodiment, MANIQA and other no-reference evaluation models are used to complete super-resolution quality evaluation. Compared with PSNR, SSIM and other no-reference evaluation models, it can fully consider the sharpness, detail preservation, texture restoration and visual perception of images, thereby providing accurate quality evaluation for each super-resolution result.

[0066] S4, based on the super-resolution quality evaluation result, output the final super-resolution result to complete image frame selection, which specifically includes the following steps:

[0067] All super-resolution quality evaluation results are converted into binary mask masks. For example, in this embodiment, one-hot mask processing method is used to normalize all super-resolution quality evaluation results, and a mask vector with the same length as the original data is further constructed, wherein only the position corresponding to the maximum score is assigned a value of 1 (such as Figure 3 In Result-2, the position is assigned a value of 1), and the remaining positions are assigned a value of 0, so as to realize explicit marking and accurate positioning of the most representative super-resolution quality evaluation result.

[0068] The super-resolution inference result corresponding to the super-resolution quality evaluation result with the Mask mask of 1 is taken as the final super-resolution result, and is output.

[0069] Thus, most current embedded AI processors (such as Rockchip RK3399, RK1126, etc.) are equipped with multiple computing cores and have a hardware basis to support parallel inference, but existing super-resolution algorithms often cannot fully exert this advantage in actual deployment. When processing real-time video streams, the system has very high requirements for the continuity of input frames, the output order, and the stability of inference timing, so that task scheduling must be highly synchronized. To ensure that each frame of image is output in sequence within a limited time, the system usually adopts a conservative inference strategy, which causes the inference load to be unable to be evenly distributed to all available cores. This scheduling mechanism limited by real-time constraints causes multiple cores or NPU sub-engines to be in a low-load or even idle state for a long time, resulting in significant waste of computing resources and weakening the performance potential of the multi-core hardware platform;

[0070] In the embodiment, the super-resolution processing unit and the super-resolution quality evaluation unit can be integrated in the same platform, the multi-Batch processing strategy is used to fully exert the parallel computing capability of the multi-core super-resolution processing unit, the processing efficiency of the super-resolution model is greatly improved, the processing process of the image can also be efficiently run in a resource-limited environment, different styles or effects of the super-resolution model can be deployed on the super-resolution processing unit according to the demand, parallel inference calculation can be completed, the demand for video memory is significantly reduced, the inference speed is accelerated, the quality of the low-resolution image is quickly improved, real-time image enhancement is realized, the robustness and real-time performance of the super-resolution processing quality are ensured;

[0071] Further, in the embodiment, the MANIQA or other reference-free evaluation model is deployed for super-resolution quality evaluation, and the super-resolution image with the best super-resolution processing effect is marked and output by using the Mask mask, so that accurate quality evaluation of the super-resolution result is realized.

[0072] Embodiment 2

[0073] The difference between the embodiment and the embodiment 1 is that, as shown in FIG. 4, the step S4 outputs the final super-resolution result based on the super-resolution quality evaluation result, to complete image frame selection, which specifically includes the following steps: Figure 6

[0074] All super-resolution quality evaluation results are converted into binary Mask masks, and the Mask mask conversion manner is the same as that of the embodiment 1, which will not be described again.

[0075] The super-resolution inference result corresponding to the super-resolution quality evaluation result with the Mask mask of 1 is taken as the super-resolution result to be selected.

[0076] ​If no new super-resolution quality evaluation result is generated within the predetermined time period, the current candidate super-resolution result is directly taken as the final super-resolution result, and outputted without updating the candidate super-resolution result;

[0077] If a new super-resolution quality evaluation result is generated within the predetermined time period, and the Mask mask of the new super-resolution quality evaluation result is 1, the super-resolution inference result corresponding to the new super-resolution quality evaluation result with the Mask mask of 1 is taken as the current candidate super-resolution result to replace the original candidate super-resolution result, so as to complete the updating of the candidate super-resolution result, and the updated candidate super-resolution result is taken as the final super-resolution result and outputted.

[0078] Therefore, in the embodiment, it can be periodically detected whether a new super-resolution quality evaluation result is generated, and if not, the current candidate super-resolution result is directly outputted to ensure the stability and consistency of the system; if so, the candidate super-resolution result is updated and then outputted, so that the change of image quality can be adaptively coped with without affecting the super-resolution inference efficiency, and the processing effect and stability of the super-resolution model are improved.

[0079] Embodiment 3:

[0080] The embodiment provides a multi-Batch super-resolution image frame selection system, which can implement the multi-Batch super-resolution image frame selection method described in embodiments 1-2, as shown in Figure 7 The multi-Batch super-resolution image frame selection system comprises:

[0081] an image segmentation unit 1 configured to divide a single-frame low-resolution image into n image blocks;

[0082] a multi-Batch inference unit 2 configured to set n batches, and each Batch is subjected to super-resolution inference on the image blocks by a corresponding super-resolution processing model, and output the super-resolution processing result of each image block;

[0083] an image splicing unit 3 configured to splice the super-resolution processing results of all the image blocks to obtain a complete high-resolution image, and output the complete high-resolution image;

[0084] N super-resolution processing units 4 configured to perform super-resolution inference on the same frame of high-resolution image, and each super-resolution processing unit outputs a corresponding super-resolution inference result;

[0085] a super-resolution quality evaluation unit 5 configured to perform super-resolution quality evaluation on each super-resolution inference result based on the super-resolution quality evaluation unit, and obtain a super-resolution quality evaluation result corresponding to each super-resolution inference result;

[0086] And, a super-resolution result output unit 6 outputs a final super-resolution result based on the super-resolution quality evaluation result to complete image frame selection.

[0087] Preferably, in the embodiment, the N super-resolution processing units 4 and the super-resolution quality evaluation unit 5 can be integrated in the same platform (such as a chip).

[0088] In summary, the present application can integrate the super-resolution processing unit and the super-resolution quality evaluation unit in the same platform, fully exert the parallel computing capability of the multi-core super-resolution processing unit through the multi-Batch processing strategy, greatly improve the processing efficiency of the super-resolution model, ensure that the image processing process can also be efficiently run in a resource-limited environment, and ensure the robustness and real-time performance of the super-resolution processing quality.

[0089] Meanwhile, the present application can realize accurate quality evaluation of the super-resolution result by deploying the MANIQA and other no-reference evaluation models for super-resolution quality evaluation, and combining the Mask mask to mark and output the super-resolution image with the best super-resolution processing effect, and can periodically detect whether a new super-resolution quality evaluation result is generated, and update the selected super-resolution result when the condition is met, so as to adaptively cope with the change of image quality without affecting the super-resolution inference efficiency, and improve the processing effect and stability of the super-resolution model.

[0090] The above description is only a preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A multi-batch super-resolution image frame selection method based on a multi-core processing unit, characterized in that: The steps include: Perform super-resolution inference on the same frame of high-resolution image, and each super-resolution processing unit outputs the corresponding super-resolution inference result; Performing super-resolution quality evaluation on each super-resolution inference result based on the super-resolution quality evaluation unit and obtaining a corresponding super-resolution quality evaluation result; The final super-resolution result is output based on the super-resolution quality assessment result to complete the image frame selection.

2. The multi-batch super-resolution image frame selection method according to claim 1, wherein: The high-resolution image is obtained by super-resolution processing of a single-frame low-resolution image, which includes the following steps: Divide a single frame low-resolution image into n image blocks; Set n batches, and each batch performs super-resolution inference on one image block, and outputs the super-resolution processing result of each image block to complete the multi-batch inference task; The super-resolution processing results of all image blocks are stitched together to obtain a complete high-resolution image.

3. The multi-batch super-resolution image frame selection method according to claim 2, wherein: Super-resolution inference for all batches of image patches is performed in parallel.

4. The multi-batch super-resolution image frame selection method according to claim 2, wherein: The n batches are all set on the same multi-batch inference unit, and each batch performs super-resolution inference on the image block through the corresponding super-resolution processing model.

5. The multi-batch super-resolution image frame selection method according to claim 1, wherein: N independent super-resolution processing units are integrated into the same platform.

6. The multi-batch super-resolution image frame selection method according to claim 1, wherein: All super-resolution processing units perform super-resolution inference on the same frame of high-resolution image in parallel.

7. The multi-batch super-resolution image frame selection method according to claim 1, wherein: At least two super-resolution processing units use different super-resolution processing models when performing super-resolution inference.

8. The multi-batch super-resolution image frame selection method according to claim 1, wherein: The final super-resolution result is output based on the super-resolution quality assessment result to complete the image frame selection, which specifically includes the following steps: Convert all super-resolution quality assessment results into binary masks; The super-resolution inference result corresponding to the super-resolution quality assessment result with Mask mask 1 is taken as the final super-resolution result and output.

9. The multi-batch super-resolution image frame selection method according to claim 1, wherein: The final super-resolution result is output based on the super-resolution quality assessment result to complete the image frame selection, which specifically includes the following steps: Convert all super-resolution quality assessment results into binary masks; The super-resolution inference result corresponding to the super-resolution quality assessment result with Mask = 1 is used as the candidate super-resolution result; If no new super-resolution quality assessment result is generated within the predetermined time period, there is no need to update the candidate super-resolution result. The current candidate super-resolution result is directly used as the final super-resolution result and output; If a new super-resolution quality assessment result with a mask of 1 is generated within a predetermined time period, the super-resolution inference result corresponding to the super-resolution quality assessment result is used as the current super-resolution result to replace the original super-resolution result to be selected, and the super-resolution result to be selected after the replacement is completed is output.

10. A multi-batch super-resolution image frame selection system, characterized by: include: N super-resolution processing units, which are used to perform super-resolution inference on the same frame of high-resolution image, and each super-resolution processing unit outputs a corresponding super-resolution inference result; a super-resolution quality evaluation unit, which performs super-resolution quality evaluation on each super-resolution inference result based on the super-resolution quality evaluation unit, and obtains a super-resolution quality evaluation result corresponding to each super-resolution inference result; And, a super-resolution result output unit outputs the final super-resolution result based on the super-resolution quality evaluation result to complete the image frame selection.

Images