Image processing system for incremental learning

Abstract

The invention discloses an incremental learning image processing system in the field of image processing, which comprises the following steps of: S1, inputting image information, namely inputting external image information into the system so as to set initial parameters for an incremental learning model in the system; s2, a preprocessing step is carried out through an incremental learning algorithm, the input image information can be marked through preprocessing of the image information, the content of the image information is classified and determined by an incremental learning algorithm model, information areas in the image are independently marked, relevance between the different information areas is set, and the content of the image information is determined; and in the preprocessing link, the color and saturation of the image can be adjusted, so that the partial image of the information area is more convenient to identify, and a plurality of learning logics can be mutually combined by adopting an incremental learning algorithm.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to an image processing system for incremental learning. Background Art

[0002] Incremental learning is a machine learning paradigm that allows the model to learn gradually as it receives new data, rather than learning the entire dataset at once. This learning style is very useful for processing large datasets, real-time data streams, or constantly changing data environments. The following are some key features and methods of incremental learning.

[0003] 1. Online Learning: In this mode, the model processes one data point or a small batch of data points at a time and updates its parameters based on these data points. This method is suitable when the data stream is continuously coming.

[0004] 2. Continuous learning: One challenge in incremental learning is to avoid catastrophic forgetting, where the model forgets knowledge of old tasks when learning new tasks. Continuous learning aims to reduce this forgetting through various strategies, such as using memory replay, regularization techniques, or innovations in neural network architectures.

[0005] 3. Lifelong learning: This refers to the process by which a model continuously learns multiple tasks or datasets throughout its lifetime. A lifelong learning model needs to be able to adapt to new tasks while maintaining knowledge of old tasks.

[0006] 4. Incremental training: In deep learning, incremental training allows the model to update its weights when it receives new data without retraining the entire network. This can be achieved through fine-tuning or using a specific incremental learning algorithm.

[0007] 5. Experience replay: This is a technique for storing previous experiences (e.g., by retaining a portion of the data or using a buffer) and reusing them when learning new tasks to help the model retain its memory for old tasks.

[0008] 6. Elastic Weight Sharing: This is a regularization method that reduces catastrophic forgetting by penalizing weight changes that have a large impact on the performance of previous tasks.

[0009] 7. Progressive Neural Networks: This approach adapts to new tasks by gradually increasing the complexity of the network while maintaining the ability to learn old tasks.

[0010] 8. Combining multi-task learning with incremental learning: In some cases, multi-task learning and incremental learning can be used simultaneously, enabling the model to learn multiple tasks at the same time and adapt to the introduction of new tasks.

[0011] Incremental learning is very valuable in practical applications, especially in environments where data is constantly updated and changing. However, it also brings some challenges, such as how to balance the learning of new and old knowledge, how to design a model architecture that can adapt to new data, and how to effectively deal with the problem of catastrophic forgetting.

[0012] Traditional image processing algorithms have the following problems:

[0013] 1. Model training takes a long time;

[0014] 2. New products cannot be used directly;

[0015] 3. When new defects are discovered, a completely new model must be trained for testing. Summary of the invention

[0016] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of the present invention to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0017] The purpose of the present invention is to address the technical problems existing in the background technology. The present invention proposes an incremental learning image processing system. The present invention can effectively improve the learning efficiency and recognition accuracy of image processing through this system. The incremental learning algorithm can be used to combine multiple learning logics and use them together. Therefore, the system can incrementally learn a recognition method to achieve multiple steps of recognition of image content, and then improve the recognition accuracy of image content through multiple recognition logics. At the same time, the method of comparing multiple recognition logics can achieve the effect of adaptive fine-tuning of recognition logic, thereby improving the recognition accuracy of image content after a period of use.

[0018] The present invention proposes an image processing system for incremental learning, comprising the following steps:

[0019] S1: image information input step, by inputting external image information into the system, so as to set initial parameters for the incremental learning model in the system;

[0020] S2 performs preprocessing steps through the incremental learning algorithm. By preprocessing the image information, the input image information can be marked. The incremental learning algorithm model classifies and determines the content categories of the image information, and separately marks the information areas in the image, and sets the correlation between different information areas. The preprocessing step can adjust the image color and saturation, thereby making the image in the information area easier to identify;

[0021] S3 uses a backtracking algorithm to perform the analysis step, inputting the pre-processed image information into the backtracking algorithm in the incremental learning algorithm for analysis. The backtracking algorithm uses a method of summarizing past data to compare the input image information with the image information analyzed in the past, and classifies the image information more accurately in the same category, and identifies and marks the information content of different information areas;

[0022] S4 uses a fixed neural network algorithm for the analysis step, inputs the preprocessed image information into the fixed neural network algorithm in the incremental learning algorithm for analysis, and the fixed neural network algorithm uses a tree-like classification recognition method to define the image multiple times through tree-like hierarchical classification, so as to finally determine the image as the minimum classification type. The fixed neural network algorithm performs independent hierarchical tree classification for different information areas, and finally integrates the results of multiple information area recognition to output the overall information content of the image information;

[0023] S5 is a step of comparing the results of the two algorithms. In this step, the recognition and classification results of the two algorithms for the same image information are compared and analyzed to obtain an algorithm that is more appropriate to the preprocessing result obtained by the incremental learning algorithm. The two algorithms are used to independently recognize each other, thereby avoiding the mutual influence of erroneous recognition results, so that erroneous recognition information will not be repeated;

[0024] S6 is a step of fine-tuning the algorithm model. This step compares the results of the two algorithms with the preprocessing results obtained by the incremental learning algorithm, and fine-tunes the parameters of the algorithm with a large difference, thereby optimizing the recognition logic unit inside the algorithm by a reverse correction method so that its output result is closer to the preprocessing result of the incremental learning algorithm;

[0025] S7 is the step for dealing with catastrophic forgetting. After the two algorithms have been fine-tuned, the image information is input into the preprocessing link again. By re-preprocessing the same image, it is possible to determine whether the learning model of the incremental learning algorithm has deviated. At the same time, the fine-tuning information is recorded during the loop processing process, so that the algorithm can be quickly reset when preprocessing changes occur, thereby dealing with the catastrophic forgetting problem.

[0026] S8 determines the learning model step. When the image information has been recognized for multiple cycles and the recognition results of the two algorithms for the image information are consistent and have a high accuracy rate, the current learning model data is locked at this time to determine the learning path from inputting the image information to outputting the recognition results of the internal content of the image information.

[0027] By adopting the above technical solution, this solution can achieve recognition learning efficiency and accuracy of image files through an incremental learning model. The incremental learning method can integrate multiple recognition logics with each other, and can add combinations based on new recognition methods discovered subsequently, thereby ensuring that the system is always in the latest recognition state.

[0028] Preferably, the preprocessing step is performed through an incremental learning algorithm, which preprocesses the image file. Through the color distribution and information area distribution of the image file, the image file can be preliminarily marked. Since the preprocessing only analyzes the image file in a large category direction, the accuracy of the preprocessing result is relatively high, and the preprocessing result for the image file is more accurate.

[0029] By adopting the above technical solution, this solution can perform basic large-scale analysis and processing on image files through pre-processing processing effects, thereby being the best reference for subsequent recognition while avoiding large deviations in analysis results.

[0030] Preferably, the step of using a backtracking algorithm to perform the analysis has an independent image buffer. By separately storing the image files identified by the algorithm, the features of the image files previously identified by the algorithm can be compared and identified during the process of the algorithm's working identification, thereby achieving content tagging of the current input image file based on the labels marked with the past image files.

[0031] By adopting the above technical solution, this solution can analyze and identify the current unknown picture through past picture records, and the recognition based on past data can effectively improve the recognition efficiency of the picture.

[0032] Preferably, the fixed neural network algorithm performs the analysis step, which will not be affected by the past image file library compared to the backtracking algorithm. The fixed tree-like classification and recognition method can gradually refine the internal classification results of the image file, and finally complete the minimum level classification of different information areas in the image file, and then through multiple minimum level classification effects, the overall recognition effect of the image file is achieved.

[0033] By adopting the above technical solution, this solution can gradually refine the image content by using a step-by-step recognition and classification method through a fixed structure of the neural network algorithm, thereby achieving the effect of independent recognition of the image. In contrast, the neural network algorithm will not be affected by past recognition data.

[0034] Preferably, the step of fine-tuning the algorithm model can adaptively adjust the system during the process of identifying different types of image files by reversely fine-tuning the algorithm, thereby improving the accuracy of targeted identification of different types of image files by multiple algorithms.

[0035] By adopting the above technical solution, this solution can judge the authenticity and accuracy of the current image content by comparing the results of the two algorithms, and evaluate two different and independent algorithms at the same time.

[0036] Preferably, the step of processing catastrophic forgetting backs up the fine-tuning data so that when a catastrophic deviation occurs in the algorithm model, the algorithm parameters can be quickly restored according to the backup data, thereby avoiding the collapse of the overall learning algorithm.

[0037] By adopting the above technical solution, the present solution can achieve the effect of self-correction of the recognition algorithm through a cyclic iteration, so that the image recognition can be more accurate as the system is used.

[0038] In summary, the present invention includes at least one of the following beneficial effects:

[0039] This system can realize the recognition of the internal content of image files, mark the image files according to the image content, and thus realize the recognition and analysis of the image content. This system adopts a combination of multiple recognition logics to improve the accuracy of image recognition, and can fine-tune the learning algorithm itself through adaptive loops, thereby realizing the effect of correcting the algorithm itself during the image recognition process, thereby improving the recognition accuracy of this system. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0041] Figure 1 A front view of an embodiment of an image processing system for incremental learning of the present invention; DETAILED DESCRIPTION

[0042] The following is combined with Figure 1 The present invention is described in further detail.

[0043] Embodiment 1

[0044] like Figure 1As shown, in this embodiment, in order to solve the existing problems, the present invention discloses an image processing system for incremental learning, including the following steps:

[0045] S1: image information input step, by inputting external image information into the system, so as to set initial parameters for the incremental learning model in the system;

[0046] S2 performs preprocessing steps through the incremental learning algorithm. By preprocessing the image information, the input image information can be marked. The incremental learning algorithm model classifies and determines the content categories of the image information, and separately marks the information areas in the image, and sets the correlation between different information areas. The preprocessing step can adjust the image color and saturation, thereby making the image in the information area easier to identify;

[0047] S3 uses a backtracking algorithm to perform the analysis step, inputting the pre-processed image information into the backtracking algorithm in the incremental learning algorithm for analysis. The backtracking algorithm uses a method of summarizing past data to compare the input image information with the image information analyzed in the past, and classifies the image information more accurately in the same category, and identifies and marks the information content of different information areas;

[0048] S4 uses a fixed neural network algorithm for the analysis step, inputs the preprocessed image information into the fixed neural network algorithm in the incremental learning algorithm for analysis, and the fixed neural network algorithm uses a tree-like classification recognition method to define the image multiple times through tree-like hierarchical classification, so as to finally determine the image as the minimum classification type. The fixed neural network algorithm performs independent hierarchical tree classification for different information areas, and finally integrates the results of multiple information area recognition to output the overall information content of the image information;

[0049] S5 is a step of comparing the results of the two algorithms. In this step, the recognition and classification results of the two algorithms for the same image information are compared and analyzed to obtain an algorithm that is more appropriate to the preprocessing result obtained by the incremental learning algorithm. The two algorithms are used to independently recognize each other, thereby avoiding the mutual influence of erroneous recognition results, so that erroneous recognition information will not be repeated;

[0050] S6 is a step of fine-tuning the algorithm model. This step compares the results of the two algorithms with the preprocessing results obtained by the incremental learning algorithm, and fine-tunes the parameters of the algorithm with a large difference, thereby optimizing the recognition logic unit inside the algorithm by a reverse correction method so that its output result is closer to the preprocessing result of the incremental learning algorithm;

[0051] S7 is the step for dealing with catastrophic forgetting. After the two algorithms have been fine-tuned, the image information is input into the preprocessing link again. By re-preprocessing the same image, it is possible to determine whether the learning model of the incremental learning algorithm has deviated. At the same time, the fine-tuning information is recorded during the loop processing process, so that the algorithm can be quickly reset when preprocessing changes occur, thereby dealing with the catastrophic forgetting problem.

[0052] S8 determines the learning model step. When the image information has been recognized for multiple cycles and the recognition results of the two algorithms for the image information are consistent and have a high accuracy rate, the current learning model data is locked at this time to determine the learning path from inputting the image information to outputting the recognition results of the internal content of the image information.

[0053] Through the incremental learning algorithm, this system can combine multiple image recognition logics and can achieve the effect of reversely improving the recognition logic by comparing the recognition results, thereby realizing the effect of the device's adaptive image recognition ability and playing the role of incremental learning.

[0054] Embodiment 2

[0055] like Figure 1 As shown, in order to solve the existing problems, based on the same concept as the above-mentioned embodiment 1, the incremental learning image processing system also includes: the preprocessing step is performed by the incremental learning algorithm, which preprocesses the image file, and can preliminarily mark the image file through the color distribution and information area distribution of the image file. Since the preprocessing only analyzes the image file in a large category direction, the accuracy of the preprocessing result is relatively high, and the preprocessing result of the image file is more accurate.

[0056] The step of using a backtracking algorithm to perform the analysis has an independent image buffer. By storing the image files identified by the algorithm separately, the features of the image files previously identified by the algorithm can be compared and identified during the process of the algorithm's working identification, thereby achieving content tagging of the current input image file based on the labels marked with the past image files.

[0057] The fixed neural network algorithm performs the analysis step, which will not be affected by the past image file library compared to the backtracking algorithm. The fixed tree-like classification and recognition method can gradually refine the internal classification results of the image file, and finally complete the minimum level classification of different information areas in the image file, and then achieve the overall recognition effect of the image file through multiple minimum level classification effects.

[0058] The step of fine-tuning the algorithm model can make adaptive adjustments in the process of identifying different types of image files in the system by reversely fine-tuning the algorithm, thereby improving the accuracy of targeted identification of different types of image files by multiple algorithms.

[0059] The catastrophic forgetting step is handled by backing up the fine-tuning data so that when a catastrophic deviation occurs in the algorithm model, the algorithm parameters can be quickly restored according to the backup data, thereby avoiding the collapse of the overall learning algorithm.

[0060] The use of a loop iteration method can achieve the effect of multiple recognition of the same photo, thereby ensuring the accuracy of the recognition effect and fine-tuning the algorithm itself through the accuracy of subsequent recognitions.

[0061] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An image processing system for incremental learning, comprising the following steps: S1: image information input step, by inputting external image information into the system, so as to set initial parameters for the incremental learning model in the system; S2 performs preprocessing steps through the incremental learning algorithm. By preprocessing the image information, the input image information can be marked. The incremental learning algorithm model classifies and determines the content categories of the image information, and separately marks the information areas in the image, and sets the correlation between different information areas. The preprocessing step can adjust the image color and saturation, thereby making the image in the information area easier to identify; S3 uses a backtracking algorithm to perform the analysis step, inputting the pre-processed image information into the backtracking algorithm in the incremental learning algorithm for analysis. The backtracking algorithm uses a method of summarizing past data to compare the input image information with the image information analyzed in the past, and classifies the image information more accurately in the same category, and identifies and marks the information content of different information areas; S4 uses a fixed neural network algorithm for the analysis step, inputs the preprocessed image information into the fixed neural network algorithm in the incremental learning algorithm for analysis, and the fixed neural network algorithm uses a tree-like classification recognition method to define the image multiple times through tree-like hierarchical classification, so as to finally determine the image as the minimum classification type. The fixed neural network algorithm performs independent hierarchical tree classification for different information areas, and finally integrates the results of multiple information area recognition to output the overall information content of the image information; S5 is a step of comparing the results of the two algorithms. In this step, the recognition and classification results of the two algorithms for the same image information are compared and analyzed to obtain an algorithm that is more appropriate to the preprocessing result obtained by the incremental learning algorithm. The two algorithms are used to independently recognize each other, thereby avoiding the mutual influence of erroneous recognition results, so that erroneous recognition information will not be repeated; S6 is a step of fine-tuning the algorithm model. This step compares the results of the two algorithms with the preprocessing results obtained by the incremental learning algorithm, and fine-tunes the parameters of the algorithm with a large difference, thereby optimizing the recognition logic unit inside the algorithm by a reverse correction method so that its output result is closer to the preprocessing result of the incremental learning algorithm; S7 is the step for dealing with catastrophic forgetting. After the two algorithms have been fine-tuned, the image information is input into the preprocessing link again. By re-preprocessing the same image, it is possible to determine whether the learning model of the incremental learning algorithm has deviated. At the same time, the fine-tuning information is recorded during the loop processing process, so that the algorithm can be quickly reset when preprocessing changes occur, thereby dealing with the catastrophic forgetting problem. S8 determines the learning model step. When the image information has been recognized for multiple cycles and the recognition results of the two algorithms for the image information are consistent and have a high accuracy rate, the current learning model data is locked at this time to determine the learning path from inputting the image information to outputting the recognition results of the internal content of the image information.

2. The image processing system for incremental learning according to claim 1, characterized in that: The preprocessing step is performed through an incremental learning algorithm. This step preprocesses the image file. Through the color distribution and information area distribution of the image file, the image file can be preliminarily marked. Since the preprocessing only analyzes the image file in a large category direction, the accuracy of the preprocessing result is relatively high, and the preprocessing result for the image file is more accurate.

3. The image processing system for incremental learning according to claim 1, characterized in that: The step of using a backtracking algorithm to perform the analysis has an independent image buffer. By storing the image files identified by the algorithm separately, the features of the image files previously identified by the algorithm can be compared and identified during the process of the algorithm's working identification, thereby achieving content tagging of the current input image file based on the labels marked with the past image files.

4. The image processing system for incremental learning according to claim 1, characterized in that: The fixed neural network algorithm performs the analysis step, which will not be affected by the past image file library compared to the backtracking algorithm. The fixed tree-like classification and recognition method can gradually refine the internal classification results of the image file, and finally complete the minimum level classification of different information areas in the image file, and then achieve the overall recognition effect of the image file through multiple minimum level classification effects.

5. The image processing system for incremental learning according to claim 1, characterized in that: The step of fine-tuning the algorithm model can make adaptive adjustments in the process of identifying different types of image files in the system by reversely fine-tuning the algorithm, thereby improving the accuracy of targeted identification of different types of image files by multiple algorithms.

6. The image processing system for incremental learning according to claim 1, characterized in that: The catastrophic forgetting step is handled by backing up the fine-tuning data so that when a catastrophic deviation occurs in the algorithm model, the algorithm parameters can be quickly restored according to the backup data, thereby avoiding the collapse of the overall learning algorithm.

Images