Image data automatic annotation processing method and system

Abstract

The invention relates to the technical field of image processing, and discloses an image data automatic labeling processing method and system, and the method comprises the steps: carrying out the preprocessing of a to-be-processed image, analyzing the content complexity of the to-be-processed image, and selecting an initial labeling model for labeling; and after labeling, obtaining a confidence coefficient score, comparing the confidence coefficient score with a qualified confidence coefficient, and judging whether to adjust a complexity threshold value or not. If adjustment is needed, reselecting a labeling model based on a new complexity threshold value, and if the models selected for the second time are different, relabeling and obtaining a second confidence coefficient; and if so, directly marking the image. And further adjusting the complexity threshold by comparing the first confidence with the second confidence. And if the models are different again, adjusting and obtaining a third confidence coefficient, and finally determining whether to output a labeling result or continue to label the image according to comparison between the third confidence coefficient and the qualified confidence coefficient. The dynamic adjustment mechanism improves the labeling precision and the system adaptability.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method and system for automatically labeling image data. Background Art

[0002] With the continuous development of artificial intelligence technology, automatic image data annotation technology has been widely used in many fields such as medical imaging, autonomous driving, and security monitoring. Automatic annotation technology can automatically identify and annotate objects, regions, and related features in large-scale image data by utilizing machine learning and deep learning algorithms, greatly improving data processing efficiency. However, existing automatic annotation methods usually have some problems, mainly reflected in the balance between annotation accuracy and efficiency, the processing of complex image content, and the adaptability and self-adaptation of annotation models.

[0003] Currently, image annotation systems face an important problem: when the image content is complex, the accuracy of the annotation model tends to decrease. This phenomenon is particularly evident when facing complex scenes or images with high detail requirements. Existing methods usually rely on a single annotation model, which may not be able to adapt to changes in image content of different complexity, resulting in low confidence in the annotation results, which in turn affects subsequent processing and decision-making.

[0004] In addition, existing image annotation systems often lack a flexible adjustment mechanism when dealing with annotation model selection. Once the annotation model is selected, it is difficult to dynamically optimize and adjust it according to the actual annotation effect. Especially when the confidence level of the annotation result is lower than the preset qualification standard, how to automatically adjust the complexity threshold and select the appropriate annotation model to improve the accuracy of the annotation result is still an urgent problem to be solved. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes an automatic image data annotation processing method and system.

[0006] On the one hand, the present invention provides an automatic image data annotation processing method, comprising: Obtain the image to be processed after preprocessing, perform content complexity analysis on each frame of the image to be processed to obtain content complexity; set a complexity threshold, select a labeling model for the first time according to the relationship between the content complexity and the complexity threshold, and label the image to be processed by using the selected initial labeling model; After the labeling is completed, a confidence score of the labeling result is obtained, recorded as a first confidence score, and according to the first confidence score and a preset qualified confidence score, it is determined whether to adjust the complexity threshold to obtain a modified complexity threshold; If the complexity threshold needs to be adjusted, a second annotation model is selected based on the relationship between the modified complexity threshold and the content complexity after the adjustment. When the second selected annotation model is different from the first selected annotation model, the image to be processed is re-annotated, and a confidence score is obtained after the annotation is completed, which is recorded as the second confidence. When the second selected annotation model is the same as the first selected annotation model, the image to be processed is marked; the first confidence is compared with the second confidence, and it is determined whether to adjust the modified complexity threshold based on the relationship between the first confidence and the second confidence. When the annotation model selected for the second time is different from the annotation model selected for the first time, the corrected complexity threshold is adjusted, and the annotation model is selected again for annotation, and a third confidence level is obtained. The third confidence level is compared with a qualified confidence level. When the third confidence level is less than the qualified confidence level, the image to be processed is marked. When the third confidence level is greater than or equal to the qualified confidence level, the annotation result is output.

[0007] Furthermore, the preprocessing includes: The image to be processed is converted into a standard format, then denoised, and the image is rotated, translated, and scaled to increase the training data for the annotation model.

[0008] Furthermore, the content complexity analysis includes: The comprehensive complexity is calculated based on edge detection, texture analysis and brightness / contrast distribution analysis; wherein, the number of edge pixels of each frame of the image to be processed is obtained based on edge detection, and the first complexity is calculated; the texture entropy of each frame of the image to be detected is calculated based on the gray level co-occurrence matrix, and the second complexity is obtained; the third complexity is obtained based on the brightness change value of each frame of the image to be detected; Weights are assigned to the first complexity, the second complexity, and the third complexity, and weighted sum is performed to obtain the content complexity of each frame of the image to be detected.

[0009] Furthermore, the first complexity, the second complexity, the third complexity and the content complexity are obtained by the following steps: The first complexity is the ratio of the number of edge pixels to the total number of pixels of each frame of the image to be processed; the second complexity is the texture entropy of each frame of the image to be processed; the third complexity is the difference between the brightness of the current frame of the image to be processed and the brightness of the previous adjacent frame to be processed; The content complexity satisfies the following relationship: ; in, is the content complexity, is the first complexity, is the second complexity, is the third complexity, , and are the first influence coefficient, the second influence coefficient and the third influence coefficient respectively. .

[0010] Further, when selecting a labeling model for the first time according to the relationship between the content complexity and the complexity threshold, it includes: setting multiple levels of the complexity threshold, and selecting a labeling model according to the relationship between several of the complexity thresholds and the content complexity; A first-level complexity threshold and a second-level complexity threshold are set in advance, and the first-level complexity threshold is less than the second-level complexity threshold; when the content complexity is less than the first-level complexity threshold, a lightweight annotation model is selected, and when the content complexity is greater than or equal to the second-level complexity threshold, a high-precision annotation model is selected, otherwise a medium-complexity annotation model is selected.

[0011] Further, when judging whether to adjust the complexity threshold to obtain a modified complexity threshold according to the first confidence level and the preset qualified confidence level, it includes: The first confidence level is compared with the qualified confidence level; when the first confidence level is greater than or equal to the qualified confidence level, the complexity threshold is not adjusted and the annotation result is output; when the first confidence level is less than the qualified confidence level, it is determined that the complexity threshold needs to be adjusted; at this time, the annotation model selected and determined for the first time is obtained; when the annotation model is a high-precision annotation model, it is determined that the adjustment condition is not met, and the image to be processed is marked; when the annotation model is not a high-precision annotation model, the complexity threshold on the right side of the content complexity is obtained, and the complexity threshold is adjusted. The modified complexity threshold obtained after adjustment satisfies the following relationship: ; in, To correct the complexity threshold, is the complexity threshold before correction, is the qualified confidence level, is the first confidence level, The unit adjustment amount.

[0012] Furthermore, the unit adjustment amount is obtained through the following relationship: ; in, is the unit adjustment amount, is the first-level complexity threshold, is the secondary complexity threshold.

[0013] Further, judging whether to adjust the modified complexity threshold according to the relationship between the first confidence level and the second confidence level includes: When the first confidence level is greater than the second confidence level, determining that the modified complexity threshold needs to be adjusted; When the first confidence level is less than or equal to the second confidence level, it is determined not to adjust the modified complexity threshold.

[0014] Further, when the modified complexity threshold is adjusted, the adjusted modified complexity threshold is recorded as the final complexity threshold. When adjusting the final complexity threshold, the complexity threshold on the right side of the content complexity is obtained, and the modified complexity threshold is adjusted to obtain the final complexity threshold after adjustment. The final complexity threshold satisfies: ; in, is the final complexity threshold, is the second confidence level.

[0015] Compared with the prior art, the present invention has the following beneficial effects: Traditional image data annotation methods often use fixed annotation models. When faced with highly complex image content, the annotation results may be inaccurate, especially when dealing with images with high detail requirements or complex backgrounds. This technical solution dynamically selects the annotation model according to the complexity of the image content, ensuring that the most appropriate annotation model can be selected from images of different complexities, thereby improving the accuracy of annotation. At the same time, by adjusting the complexity threshold, the annotation model can be better adapted to complex scenes, enhancing the system's adaptive ability.

[0016] Through multi-level analysis of the complexity of image content (including edges, textures, brightness and other information), this solution can accurately evaluate the complexity of each frame of the image and select lightweight, medium-complexity or high-precision annotation models based on this complexity. This on-demand model selection method effectively avoids the use of inappropriate models, reduces the error rate when processing complex scenes, and improves computational efficiency.

[0017] During the annotation process, if the confidence of the preliminary annotation results is low, the existing system often cannot automatically adjust to improve accuracy. By introducing a mechanism to dynamically adjust the complexity threshold, this solution can automatically adjust when the annotation results do not meet expectations. By reselecting the annotation model and judging whether the complexity threshold needs to be further adjusted based on the new annotation results, this self-optimization mechanism ensures that the image annotation system can flexibly respond to different situations and avoid a decrease in annotation accuracy.

[0018] The "dynamic annotation model selection" and "complexity threshold adjustment" in this solution can automatically complete the annotation task in most cases, and automatically optimize through the adjustment mechanism when the annotation does not meet expectations. This automated mechanism can effectively reduce manual intervention. At the same time, when the confidence level is low, the system can also improve the quality of the annotation results through manual assistance. This combination of automation and manual collaboration has greatly improved the work efficiency of the annotation system and reduced labor costs.

[0019] This technical solution has high scalability and can adapt to the image data annotation needs in different fields. For example, in medical images, the complexity and precision of the images are high. This solution can select the appropriate annotation model according to the specific scene. It is also applicable in security monitoring, satellite remote sensing and other fields. The system can automatically adjust the annotation model for image data of different complexities, thus having cross-domain applicability.

[0020] This technical solution can not only improve the accuracy of annotation, but also enhance the quality control ability of the system by evaluating and adjusting the confidence of the annotation results multiple times. At each step of the annotation process, the system tracks the confidence and adjusts the complexity threshold and annotation model according to the actual annotation effect. Finally, when the confidence reaches the qualified standard, the system outputs high-quality annotation results. This can greatly reduce the negative impact of low-quality annotation and ensure the high quality of the annotation data.

[0021] When faced with complex images with rich details, this solution can accurately assess the complexity and flexibly select a high-precision annotation model. During the annotation process, the annotation strategy and complexity threshold can be dynamically adjusted to ensure the processing effect of highly complex images. This method can effectively deal with image data in complex scenes, and provide accurate annotation information for subsequent processing, improving the efficiency of the entire image analysis system.

[0022] On the other hand, the present invention also proposes an image data automatic annotation processing system, comprising: The first module is configured to obtain the image to be processed after preprocessing, perform content complexity analysis on each frame of the image to be processed to obtain content complexity; set a complexity threshold, select a labeling model for the first time according to the relationship between the content complexity and the complexity threshold, and label the image to be processed by using the selected initial labeling model; The second module is configured to obtain a confidence score of the annotation result after the annotation is completed, recorded as a first confidence score, and determine whether to adjust the complexity threshold to obtain a modified complexity threshold according to the first confidence score and a preset qualified confidence score; The third module is configured to, if it is necessary to adjust the complexity threshold, select a labeling model for the second time according to the relationship between the modified complexity threshold after the adjustment and the content complexity; when the labeling model selected for the second time is different from the labeling model selected for the first time, re-label the image to be processed, and obtain a confidence score after the labeling is completed, recorded as a second confidence; when the labeling model selected for the second time is the same as the labeling model selected for the first time, mark the image to be processed; compare the first confidence with the second confidence, and determine whether to adjust the modified complexity threshold according to the relationship between the first confidence and the second confidence; The fourth module is configured to adjust the corrected complexity threshold when the annotation model selected for the second time is different from the annotation model selected for the first time, and select the annotation model again for annotation, obtain a third confidence level, compare the third confidence level with the qualified confidence level, and mark the image to be processed when the third confidence level is less than the qualified confidence level, and output the annotation result when the third confidence level is greater than or equal to the qualified confidence level.

[0023] It is understandable that the image data automatic annotation processing system and method of the present invention have the same beneficial effects, which will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings: Figure 1 The present invention provides a flowchart of an automatic image data annotation processing method.

[0025] Figure 2 A functional framework diagram of an image data automatic annotation processing system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The exemplary embodiments disclosed in the present application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features described in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0027] See also Figure 1As shown, an embodiment of the present invention provides a method for automatically annotating image data, comprising: S1: Obtain the image to be processed after preprocessing, perform content complexity analysis on each frame of the image to be processed, and obtain the content complexity; set a complexity threshold, select a labeling model for the first time according to the relationship between the content complexity and the complexity threshold, and label the image to be processed by using the selected initial labeling model; S2: After the labeling is completed, the confidence score of the labeling result is obtained, recorded as the first confidence, and whether to adjust the complexity threshold to obtain the modified complexity threshold is determined according to the first confidence and the preset qualified confidence; S3: If the complexity threshold needs to be adjusted, a second annotation model is selected based on the relationship between the modified complexity threshold after adjustment and the content complexity. When the second selected annotation model is different from the first selected annotation model, the image to be processed is re-annotated, and a confidence score is obtained after the annotation is completed, which is recorded as the second confidence. When the second selected annotation model is the same as the first selected annotation model, the image to be processed is marked; the first confidence is compared with the second confidence, and it is determined whether to adjust the modified complexity threshold based on the relationship between the first confidence and the second confidence. S4: When the annotation model selected for the second time is different from the annotation model selected for the first time, the correction complexity threshold is adjusted, and the annotation model is selected again for annotation, and the third confidence is obtained. The third confidence is compared with the qualified confidence. When the third confidence is less than the qualified confidence, the image to be processed is marked. When the third confidence is greater than or equal to the qualified confidence, the annotation result is output.

[0028] It should be noted that the images to be tested come from drone aerial photography, satellite remote sensing, medical imaging, autonomous driving data, industrial inspection and other image data.

[0029] "Secondarily selecting the annotation model based on the relationship between the adjusted revised complexity threshold and the content complexity" is the same as the method of selecting the annotation model for the first time. The difference is that the adjusted revised complexity threshold replaces the original complexity threshold, and then any one of the lightweight annotation model, medium complexity annotation model and high-precision annotation model is selected.

[0030] It should be noted that this technical solution has the following beneficial effects: Adaptive annotation model selection: Dynamically match the appropriate annotation model through content complexity analysis to improve annotation efficiency and accuracy.

[0031] Optimize the utilization of computing resources: Lightweight images use low-computation models, and high-complexity images use high-precision models to achieve reasonable allocation of computing resources.

[0032] Dynamically adjust the complexity threshold: Based on annotation confidence feedback, adjust the complexity threshold to adapt to different image features and improve annotation stability.

[0033] Multi-round confidence verification mechanism: Through multiple confidence comparisons and model adjustments, the accuracy of the final annotation results is ensured, erroneous annotations are reduced, and the quality of automated annotation is improved.

[0034] Reduce manual intervention: Reduce the need for manual review through intelligent adjustment mechanisms, and improve the autonomy and reliability of automatic labeling.

[0035] Wide adaptability: Applicable to image data of different content complexity, improving the generalization ability in various application scenarios during annotation.

[0036] In some embodiments of the present application, preprocessing includes: The image to be processed is converted into a standard format, then denoised, and the image is rotated, translated, and scaled to increase the training data for the annotation model.

[0037] It should be noted that different data sources may use different image formats and encoding methods. Converting to a unified standard format (such as JPEG, PNG or TIFF) can improve processing compatibility and ensure that the annotation model can correctly interpret the input data. Images may be affected by noise, such as sensor noise, compression artifacts or environmental interference. Reducing noise through filtering (such as Gaussian filtering, median filtering or wavelet denoising) can help improve the accuracy of feature extraction such as edge detection and texture analysis.

[0038] In some embodiments of the present application, content complexity analysis includes: The comprehensive complexity is calculated based on edge detection, texture analysis and brightness / contrast distribution analysis; wherein, the number of edge pixels of each frame of the image to be processed is obtained based on edge detection, and the first complexity is calculated; the texture entropy of each frame of the image to be detected is calculated based on the gray level co-occurrence matrix, and the second complexity is obtained; the third complexity is obtained based on the brightness change value of each frame of the image to be detected; Weights are assigned to the first complexity, the second complexity, and the third complexity, and the weighted sum is taken to obtain the content complexity of each frame of the image to be detected.

[0039] It should be noted that In some embodiments of the present application, the first complexity, the second complexity, the third complexity and the content complexity are obtained by the following steps: The first complexity is the ratio of the number of edge pixels to the total number of pixels in each frame of the image to be processed; the second complexity is the texture entropy of each frame of the image to be processed; the third complexity is the difference between the brightness of the current frame of the image to be processed and the brightness of the previous adjacent frame to be processed; Content complexity satisfies the following relationship: ; in, is the content complexity, is the first complexity, is the second complexity, is the third complexity, , and are the first influence coefficient, the second influence coefficient and the third influence coefficient respectively. .

[0040] It should be noted that traditional image analysis usually relies on only a single feature, such as edge information or brightness distribution, while this method combines edge detection, texture analysis and brightness changes to calculate the comprehensive complexity, so that the content characteristics of the image can be more comprehensively quantitatively evaluated. This multi-dimensional calculation method makes the image complexity assessment more accurate, which helps to select the appropriate model in the annotation process and improve the annotation efficiency and accuracy.

[0041] The first complexity is calculated by edge detection, which can identify the structural information of the image and is suitable for scenes with clear target boundaries; the texture entropy calculated by gray-level co-occurrence matrix is ​​used as the second complexity, which is suitable for images with rich textures or more details; the third complexity is calculated by brightness change, which is helpful for analyzing images with large changes in lighting conditions. By integrating the three complexities, this method can be applied to different types of images and improve its adaptability to complex environments.

[0042] By setting weights for the three complexities and taking weighted sums, this method can adjust the weight parameters according to different task requirements. For example, when processing industrial inspection images, the weight of edge detection can be increased, while in remote sensing image processing, the influence of texture analysis can be enhanced. This flexibility enables the method to maintain good generalization capabilities in multiple fields.

[0043] Since the content complexity of different images varies, directly using a unified annotation model may lead to a decrease in annotation accuracy. This method uses a comprehensive complexity calculation to ensure that a lightweight model is used for simple images and a high-precision model is used for complex images. This avoids the waste of computing resources and reduces the misjudgment of low-complexity images due to overfitting, as well as the omission of labels for high-complexity images due to insufficient model capabilities, thereby significantly improving the overall annotation accuracy.

[0044] Feature analysis such as brightness change and texture entropy can effectively reduce the impact of interference factors such as lighting change and noise, reduce misjudgment caused by environmental factors, and improve the stability of complex calculations. At the same time, this method can adapt to image inputs of different resolutions and qualities, maintain consistent performance under different shooting conditions, and improve the reliability of the image data annotation process.

[0045] By automatically calculating the complexity of the image and dynamically adjusting the selection of the annotation model, this method reduces manual intervention and improves the automation of image data processing. This is particularly important for the annotation task of large-scale data sets, which can effectively reduce labor costs while ensuring the consistency of annotation quality.

[0046] In some embodiments of the present application, when selecting a labeling model for the first time according to the relationship between content complexity and complexity threshold, it includes: setting multiple levels of complexity thresholds, and selecting a labeling model according to the relationship between several complexity thresholds and content complexity; The first-level complexity threshold and the second-level complexity threshold are set in advance, and the first-level complexity threshold is smaller than the second-level complexity threshold; when the content complexity is smaller than the first-level complexity threshold, a lightweight annotation model is selected; when the content complexity is greater than or equal to the second-level complexity threshold, a high-precision annotation model is selected; otherwise, a medium-complexity annotation model is selected.

[0047] It should be noted that the annotation models include: lightweight annotation models, medium-complexity annotation models, and high-precision annotation models; Each annotation model includes at least one model capable of completing automatic annotation processing of image data. In this embodiment, when the current model is determined to be a level model, it is divided according to the computational overhead of the model.

[0048] It should be noted that by setting multiple complexity thresholds, simple images use lightweight models to reduce computational overhead and improve processing speed, while complex images use high-precision models to ensure annotation quality, thereby improving overall computational efficiency while ensuring annotation accuracy.

[0049] Multi-level complexity thresholds are used to distinguish image types, so that the annotation model matches the actual complexity of the image, avoiding overfitting caused by the use of overly complex models for low-complexity images, and avoiding mislabeling of high-complexity images due to insufficient model capabilities, thereby improving the overall annotation accuracy.

[0050] This method is applicable to different types of image scenes and can adjust the complexity threshold according to task requirements to adapt it to different data distributions, improve its versatility and robustness, and enhance the model's adaptability to diverse data.

[0051] By setting multiple complexity thresholds, annotation model selection can be performed automatically without human intervention, thereby reducing labor costs, improving the intelligence level of the image annotation process, and facilitating the efficient execution of large-scale automatic annotation tasks.

[0052] In some embodiments of the present application, when determining whether to adjust the complexity threshold to obtain a modified complexity threshold according to the first confidence level and the preset qualified confidence level, the method includes: The first confidence level is compared with the qualified confidence level; when the first confidence level is greater than or equal to the qualified confidence level, the complexity threshold is not adjusted and the annotation result is output; when the first confidence level is less than the qualified confidence level, it is determined that the complexity threshold needs to be adjusted; at this time, the annotation model selected and determined for the first time is obtained; when the annotation model is a high-precision annotation model, it is determined that it does not meet the adjustment conditions and the image to be processed is marked; when the annotation model is not a high-precision annotation model, the complexity threshold on the right side of the content complexity is obtained, and the complexity threshold is adjusted. The modified complexity threshold obtained after adjustment satisfies the following relationship: ; in, To correct the complexity threshold, is the complexity threshold before correction, is the qualified confidence level, is the first confidence level, The unit adjustment amount.

[0053] It should be noted that the complexity threshold is adjusted dynamically through the confidence feedback mechanism to make the model selection more accurate, avoid unreasonable annotation model selection caused by fixed thresholds, and ensure more accurate annotation results. The complexity threshold is adjusted only when the confidence is lower than the qualified standard, and the model is reselected when the conditions are met, thereby reducing unnecessary model switching, reducing computational overhead, and improving processing efficiency. This method adaptively adjusts the complexity threshold according to the actual annotation confidence of the image, can flexibly respond to different image types, improve the adaptability to diversified data, and enhance robustness in complex environments. The complexity threshold is adjusted only when the model is not high-precision and the confidence is low, and the unit adjustment amount is used to control the change range to ensure that the adjustment process is stable and controllable, and avoid instability caused by excessive adjustment. By automatically determining whether the complexity threshold needs to be adjusted and performing reasonable dynamic optimization, this method can reduce manual intervention, improve the intelligence and automation level of the annotation process, and make it more suitable for large-scale image data processing tasks.

[0054] In some embodiments of the present application, the unit adjustment amount is obtained by the following relationship: ; in, is the unit adjustment amount, is the first-level complexity threshold, is the secondary complexity threshold.

[0055] It should be noted that by setting the unit adjustment amount k to 1 / 3 of the difference between the two complexity thresholds, smooth adjustment is achieved to avoid the adjustment range being too large or too small, thereby improving the accuracy of complexity threshold control.

[0056] A hierarchical adjustment strategy is adopted, adjusting only 1 / 3 of the range each time to ensure that the threshold changes moderately and avoid instability caused by sudden changes in parameters.

[0057] By adjusting the value evenly in steps rather than jumping to extreme values ​​all at once, the optimization convergence speed can be accelerated, making the selection process of the annotation model more efficient.

[0058] The calculation method based on the primary and secondary complexity thresholds is adopted to make the adjustment method applicable to different image complexity situations, thereby improving the versatility and adaptability.

[0059] Only simple numerical calculations are needed to determine the unit adjustment amount, which reduces unnecessary complex calculations, reduces computing resource consumption, and improves processing efficiency.

[0060] In some embodiments of the present application, judging whether to adjust the modified complexity threshold according to the relationship between the first confidence level and the second confidence level includes: When the first confidence level is greater than the second confidence level, it is determined that the correction complexity threshold needs to be adjusted; When the first confidence level is less than or equal to the second confidence level, it is determined not to adjust the modified complexity threshold.

[0061] It should be noted that by comparing the first confidence level and the second confidence level, it is ensured that the revised complexity threshold does not reduce the reliability of the annotation result, thereby improving the accuracy of the final annotation.

[0062] The complexity threshold is adjusted only when the first confidence level is greater than the second confidence level, thereby reducing invalid adjustment operations and improving stability and computational efficiency.

[0063] By dynamically adjusting and correcting the complexity threshold, the selection of the annotation model can be more adapted to the actual complexity of the image, thereby improving the model's adaptability and reducing the misjudgment rate.

[0064] Adjustments are only made when the annotation quality fails to improve, to avoid over-calculation, reduce computing resource consumption, and improve operational efficiency.

[0065] By real-time monitoring of confidence changes and dynamically adjusting the complexity threshold, it can adapt to different image scenes and improve the versatility of annotation.

[0066] In some embodiments of the present application, when the modified complexity threshold is adjusted, the adjusted modified complexity threshold is recorded as the final complexity threshold. When the final complexity threshold is adjusted, the complexity threshold on the right side of the content complexity is obtained, and the modified complexity threshold is adjusted to obtain the final complexity threshold after adjustment. The final complexity threshold satisfies: ; in, is the final complexity threshold, is the second confidence level.

[0067] It should be noted that by calculating the final complexity threshold, the adjusted complexity threshold is more in line with the actual situation of the image content, thereby improving the applicability of the annotation model. The complexity threshold is adjusted in combination with the confidence P2, so that the final threshold can better match the annotation reliability of the model and reduce the occurrence of low-confidence annotation results. The complexity threshold on the right side of the image content complexity is dynamically adjusted to ensure that changes in image complexity can be reasonably responded to and improve the adaptive ability of annotation. A reasonable adjustment formula is used to ensure that the complexity threshold adjustment is not too large or too small, to prevent the stability of the annotation model from being affected by excessive parameter changes. Through targeted complexity threshold adjustment, unnecessary calculations can be reduced.

[0068] See also Figure 2 As shown, the embodiment of the present invention further proposes an image data automatic annotation processing system, including: The first module is configured to obtain the image to be processed after preprocessing, perform content complexity analysis on each frame of the image to be processed, and obtain content complexity; set a complexity threshold, select a labeling model for the first time according to the relationship between the content complexity and the complexity threshold, and label the image to be processed by using the selected initial labeling model; The second module is configured to obtain a confidence score of the annotation result after the annotation is completed, recorded as a first confidence score, and determine whether to adjust the complexity threshold to obtain a modified complexity threshold according to the first confidence score and a preset qualified confidence score; The third module is configured to, if the complexity threshold needs to be adjusted, select a labeling model for the second time according to the relationship between the modified complexity threshold after the adjustment and the content complexity; when the labeling model selected for the second time is different from the labeling model selected for the first time, re-label the image to be processed, and obtain a confidence score after the labeling is completed, which is recorded as the second confidence; when the labeling model selected for the second time is the same as the labeling model selected for the first time, mark the image to be processed; compare the first confidence with the second confidence, and determine whether to adjust the modified complexity threshold according to the relationship between the first confidence and the second confidence; The fourth module is configured to adjust the correction complexity threshold when the annotation model selected for the second time is different from the annotation model selected for the first time, and select the annotation model again for annotation, obtain the third confidence level, compare the third confidence level with the qualified confidence level, and mark the image to be processed when the third confidence level is less than the qualified confidence level, and output the annotation result when the third confidence level is greater than or equal to the qualified confidence level.

[0069] It should be noted that the beneficial effects of the above technical solution include the following aspects: By analyzing the complexity of the content, a suitable annotation model is selected to make the initial annotation results more consistent with the characteristics of the image, reducing mislabeling and missing labels. Multiple rounds of confidence verification are used to ensure that the final output annotation results have high confidence and improve the reliability of the overall annotation. The complexity threshold is set and dynamically adjusted based on the actual annotation confidence to avoid the problem of insufficient generalization caused by the fixed threshold. In the annotation process, different levels of models are selected according to the complexity of the image to ensure that neither over-calculation nor errors are caused by insufficient model capabilities. Combined with multiple adjustment mechanisms, the complexity threshold is gradually optimized by comparing confidence, so that the system can adapt to different types of image data. A step-by-step screening mechanism is adopted to give priority to the use of lightweight models to process simple images, reduce the frequency of calling high-precision models, and thus optimize computing resources. Secondary or tertiary annotation is performed only when necessary to avoid invalid repeated calculations and improve the operating efficiency of the entire system. By gradually adjusting the complexity threshold, the calculation instability caused by large-scale parameter adjustments is reduced, and the system response speed is improved. Through multiple rounds of confidence evaluation, adaptive optimization of the annotation process is achieved, manual inspection and intervention are reduced, and the degree of automation of image data processing is improved. The intelligent threshold adjustment strategy enables the system to automatically adjust according to the data characteristics, without the need to manually set the complexity range of each image. Through content complexity analysis, it can adapt to different categories of images and improve the versatility of the annotation system. The multi-level annotation model is adopted to ensure that the precision of different images matches the appropriate annotation strategy to avoid the problem of over-annotation or under-annotation. By comprehensively evaluating the first confidence, second confidence and third confidence, the final output annotation results have higher reliability and consistency. The multi-level confidence comparison is used to prevent the accumulation of errors caused by the wrong selection of the initial annotation model. By gradually adjusting the complexity threshold, it is ensured that the system can self-correct, reduce the result deviation caused by a single annotation error, and improve the stability of the overall system.

[0070] This solution improves the accuracy, efficiency and intelligence of automatic image annotation through strategies such as multi-level preprocessing, dynamic model selection, confidence assessment and adaptive adjustment of complexity threshold. It not only reduces the computing cost, but also enhances the adaptability of the system. It is suitable for a variety of image processing scenarios and can be widely used in medical imaging, remote sensing data, intelligent monitoring and other fields.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for automatically labeling image data, characterized in that: include: Obtain the image to be processed after preprocessing, perform content complexity analysis on each frame of the image to be processed, and obtain content complexity; Setting a complexity threshold, selecting a labeling model for the first time according to the relationship between the content complexity and the complexity threshold, and labeling the image to be processed by using the selected initial labeling model; After the labeling is completed, a confidence score of the labeling result is obtained, recorded as a first confidence score, and according to the first confidence score and a preset qualified confidence score, it is determined whether to adjust the complexity threshold to obtain a modified complexity threshold; If the complexity threshold needs to be adjusted, a second annotation model is selected based on the relationship between the modified complexity threshold and the content complexity after the adjustment. When the second selected annotation model is different from the first selected annotation model, the image to be processed is re-annotated, and a confidence score is obtained after the annotation is completed, which is recorded as the second confidence score. When the second selected annotation model is the same as the first selected annotation model, the image to be processed is marked; Comparing the first confidence level with the second confidence level, and determining whether to adjust the modified complexity threshold according to a relationship between the first confidence level and the second confidence level; When the annotation model selected for the second time is different from the annotation model selected for the first time, the corrected complexity threshold is adjusted, and the annotation model is selected again for annotation, and a third confidence level is obtained. The third confidence level is compared with a qualified confidence level. When the third confidence level is less than the qualified confidence level, the image to be processed is marked. When the third confidence level is greater than or equal to the qualified confidence level, the annotation result is output.

2. The method for automatic image data annotation processing according to claim 1, characterized in that: The pre-processing comprises: The image to be processed is converted into a standard format and then denoised. The image is rotated, translated, and scaled to increase the training data for the annotation model.

3. The method for automatic image data annotation processing according to claim 2, characterized in that: The content complexity analysis includes: The comprehensive complexity is calculated based on edge detection, texture analysis and brightness / contrast distribution analysis; wherein, the number of edge pixels of each frame of the image to be processed is obtained based on edge detection, and the first complexity is calculated; the texture entropy of each frame of the image to be detected is calculated based on the gray level co-occurrence matrix, and the second complexity is obtained; the third complexity is obtained based on the brightness change value of each frame of the image to be detected; Weights are assigned to the first complexity, the second complexity, and the third complexity, and weighted sum is performed to obtain the content complexity of each frame of the image to be detected.

4. The method for automatic image data annotation processing according to claim 3, characterized in that: The first complexity, the second complexity, the third complexity and the content complexity are obtained by the following steps: The first complexity is the ratio of the number of edge pixels to the total number of pixels of each frame of the image to be processed; the second complexity is the texture entropy of each frame of the image to be processed; the third complexity is the difference between the brightness of the current frame of the image to be processed and the brightness of the previous adjacent frame to be processed; The content complexity satisfies the following relationship: ； in, is the content complexity, is the first complexity, is the second complexity, is the third complexity, , and are the first influence coefficient, the second influence coefficient and the third influence coefficient respectively. .

5. The method for automatic image data annotation processing according to claim 4, characterized in that: When selecting a labeling model for the first time according to the relationship between the content complexity and the complexity threshold, it includes: setting multiple levels of the complexity threshold, and selecting a labeling model according to the relationship between several of the complexity thresholds and the content complexity; A first-level complexity threshold and a second-level complexity threshold are set in advance, and the first-level complexity threshold is less than the second-level complexity threshold; when the content complexity is less than the first-level complexity threshold, a lightweight annotation model is selected, and when the content complexity is greater than or equal to the second-level complexity threshold, a high-precision annotation model is selected, otherwise a medium-complexity annotation model is selected.

6. The method for automatic image data annotation processing according to claim 5, characterized in that: When judging whether to adjust the complexity threshold to obtain a modified complexity threshold according to the first confidence level and the preset qualified confidence level, the method includes: The first confidence level is compared with the qualified confidence level; when the first confidence level is greater than or equal to the qualified confidence level, the complexity threshold is not adjusted and the annotation result is output; when the first confidence level is less than the qualified confidence level, it is determined that the complexity threshold needs to be adjusted; at this time, the annotation model selected and determined for the first time is obtained; when the annotation model is a high-precision annotation model, it is determined that the adjustment condition is not met, and the image to be processed is marked; when the annotation model is not a high-precision annotation model, the complexity threshold on the right side of the content complexity is obtained, and the complexity threshold is adjusted. The modified complexity threshold obtained after adjustment satisfies the following relationship: ； in, To correct the complexity threshold, is the complexity threshold before correction, is the qualified confidence level, is the first confidence level, The unit adjustment amount.

7. The method for automatic image data annotation processing according to claim 6, characterized in that: The unit adjustment amount is obtained through the following relationship: ； in, is the unit adjustment amount, is the first-level complexity threshold, is the secondary complexity threshold.

8. The method for automatic image data annotation processing according to claim 7, characterized in that: When judging whether to adjust the modified complexity threshold according to the relationship between the first confidence level and the second confidence level, the method includes: When the first confidence level is greater than the second confidence level, determining that the modified complexity threshold needs to be adjusted; When the first confidence level is less than or equal to the second confidence level, it is determined not to adjust the modified complexity threshold.

9. The method for automatic image data annotation processing according to claim 8, characterized in that: When the modified complexity threshold is adjusted, the adjusted modified complexity threshold is recorded as the final complexity threshold. When adjusting the final complexity threshold, the complexity threshold on the right side of the content complexity is obtained, and the modified complexity threshold is adjusted to obtain the final complexity threshold after adjustment. The final complexity threshold satisfies: ； in, is the final complexity threshold, is the second confidence level.

10. An automatic image data annotation processing system, used to implement the automatic image data annotation processing method according to any one of claims 1 to 9, characterized in that: include: The first module is configured to obtain the image to be processed after preprocessing, and perform content complexity analysis on each frame of the image to be processed to obtain content complexity; Setting a complexity threshold, selecting a labeling model for the first time according to the relationship between the content complexity and the complexity threshold, and labeling the image to be processed by using the selected initial labeling model; The second module is configured to obtain a confidence score of the annotation result after the annotation is completed, recorded as a first confidence score, and determine whether to adjust the complexity threshold to obtain a modified complexity threshold according to the first confidence score and a preset qualified confidence score; The third module is configured to, if the complexity threshold needs to be adjusted, select a labeling model for the second time according to the relationship between the modified complexity threshold after the adjustment and the content complexity, and when the labeling model selected for the second time is different from the labeling model selected for the first time, re-label the image to be processed, and obtain a confidence score after the labeling is completed, which is recorded as a second confidence score, and when the labeling model selected for the second time is the same as the labeling model selected for the first time, mark the image to be processed; Comparing the first confidence level with the second confidence level, and determining whether to adjust the modified complexity threshold according to a relationship between the first confidence level and the second confidence level; The fourth module is configured to adjust the corrected complexity threshold when the annotation model selected for the second time is different from the annotation model selected for the first time, and select the annotation model again for annotation, obtain a third confidence level, compare the third confidence level with the qualified confidence level, and mark the image to be processed when the third confidence level is less than the qualified confidence level, and output the annotation result when the third confidence level is greater than or equal to the qualified confidence level.

Images