CN118334663A

CN118334663A - One-stop artificial intelligent image processing model construction method and device

Info

Publication number: CN118334663A
Application number: CN202410760236.9A
Authority: CN
Inventors: 郑东; 殷海兵; 刘浩; 赵五岳; 赵拯; 徐宇杰; 胡榛旸
Original assignee: Hangzhou Yufan Intelligent Technology Co ltd; Hangzhou Dianzi University
Current assignee: Hangzhou Yufan Intelligent Technology Co ltd; Hangzhou Dianzi University
Priority date: 2024-06-13
Filing date: 2024-06-13
Publication date: 2024-07-12
Anticipated expiration: 2044-06-13
Also published as: CN118334663B

Abstract

The embodiment of the application provides a method and a device for constructing a one-stop artificial intelligent image processing model, wherein the method comprises the following steps: creating an image annotation task, determining the structure of a corresponding image processing model, and training the image processing model according to the image data stored in the standardized label to obtain the image processing model; determining a corresponding pre-processing interface, an inference interface and a post-processing interface of the target deployment hardware according to a model definition specification corresponding to the image processing model, inputting image data stored by a standardized tag processed by the pre-processing interface into the image processing model through the inference interface, processing model output of the image processing model through the post-processing interface, determining corresponding model precision according to the model output, and deploying the image processing model to the target deployment hardware when the model precision meets a preset precision condition; the application can realize the integrated integration of the whole processes of data labeling, model training, model conversion, model deployment and the like, improves the development efficiency, reduces the development cost and ensures the high consistency of the training precision and the deployment precision.

Description

One-stop artificial intelligent image processing model construction method and device

Technical Field

The application relates to the field of artificial intelligence, in particular to a one-stop artificial intelligence image processing model construction method and device.

Background

Artificial intelligence technology has been rapidly developed in recent years, and significant progress has been made in the fields of computer vision, natural language processing, speech recognition, and the like. A complete artificial intelligence model development typically includes multiple steps of data collection and annotation, model training and testing, model conversion and deployment, and the like. However, existing artificial intelligence model building processes present a number of problems and challenges.

Firstly, data annotation is the basis of artificial intelligent model development, but the traditional data annotation has lower working efficiency and has the problem of non-uniform tag data format. Under the normal condition, the data marking needs to be carried out manually, the flow is complex, and the efficiency is low. Meanwhile, the labeling formats used by different projects or teams are different, and a plurality of inconveniences are brought to subsequent model training and deployment.

Secondly, model training and testing are also key links in artificial intelligence model development. Currently, developers often need to switch between multiple different platforms and tools, such as model training using Pytorch or TensorFlow, and model conversion and quantification by ONNX or TensorRT. The development mode of the fragmentation is not only inefficient, but also prone to human error. In addition, the existing model training mode often cannot fully consider the hardware performance and deployment requirements of an actual deployment environment, so that the trained model cannot fully exert performance in actual deployment.

Furthermore, model conversion and quantization are also key links in artificial intelligence model development. Different hardware platforms have differences in operator support, precision requirements and the like of the model, and the model conversion and quantization are required to be performed pertinently. This typically requires extensive hardware knowledge and programming experience by the developer to complete the conversion process. In reality, however, most algorithm engineers are not hardware specialists, and it is difficult to efficiently perform model conversion and quantization work.

Finally, there are also problems with the model deployment phase. Even through the foregoing model conversion and quantization, the trained model may still have a reduced accuracy in actual deployment. This is because the model differs between the training environment and the deployment environment in that detailed information such as preprocessing of the input data, post-processing logic, etc. cannot be aligned exactly. The training accuracy cannot fully correspond to the actual deployment accuracy, which brings uncertainty to the application landing.

In summary, the existing artificial intelligent model development process has the problems of low data labeling efficiency, low model training efficiency, complex model conversion, unaligned deployment precision and the like. These problems not only increase the complexity and cost of artificial intelligence model development, but also greatly limit the popularity and popularization of artificial intelligence technology in practical applications.

Disclosure of Invention

Aiming at the problems in the prior art, the application provides a one-stop artificial intelligent image processing model construction method and device, which can realize the integrated integration of the whole processes of data labeling, model training, model conversion, model deployment and the like, improve the development efficiency, reduce the development cost and ensure the high consistency of the training precision and the deployment precision.

In order to solve at least one of the problems, the application provides the following technical scheme:

in a first aspect, the present application provides a method for constructing a one-stop artificial intelligence image processing model, including:

Creating an image annotation task, determining a corresponding pre-annotation model according to the image annotation task, performing pre-annotation processing on preset image data according to the pre-annotation model, and storing the image data subjected to the pre-annotation processing in a corresponding standardized tag according to a data type;

receiving target deployment hardware and target model running speeds selected by a user, determining the structure of a corresponding image processing model according to the numerical comparison relation between the pre-evaluation speed of each preset image processing model on the target deployment hardware and the target model running speed, and training the image processing model according to image data stored by the standardized label to obtain the image processing model;

Converting the image processing model into a corresponding executable format according to the specific type of the target deployment hardware, and quantizing floating point parameters of the image processing model into integers or fixed point numbers to obtain an image processing model subjected to format conversion and parameter quantization;

Determining a pre-processing interface, an inference interface and a post-processing interface corresponding to the target deployment hardware according to a model definition specification corresponding to the image processing model, inputting image data stored by the standardized tag after being processed by the pre-processing interface to the image processing model through the inference interface, processing model output of the image processing model through the post-processing interface, determining corresponding model precision according to the model output, and deploying the image processing model to the target deployment hardware when the model precision meets a preset precision condition.

Further, the determining a corresponding pre-labeling model according to the image labeling task, and performing pre-labeling processing on preset image data according to the pre-labeling model includes:

determining a corresponding pre-labeling model according to the task type of the image labeling task and the data characteristics of preset image data;

and inputting the preset image data into the pre-labeling model for pre-labeling treatment to obtain pre-labeled image data.

Further, the determining a corresponding pre-labeling model according to the task type of the image labeling task and the data characteristics of the preset image data includes:

determining task types of the image labeling task, and determining image recognition emphasis features according to the task types, wherein the task types comprise an image classification task, an image detection task and an image segmentation task, and the image recognition emphasis features comprise emphasis global features and emphasis local features;

And extracting key features of preset image data to obtain corresponding image features, and matching corresponding pre-labeling models from a preset model library according to the image features and the task types.

Further, the matching the corresponding pre-labeling model from the pre-setting model library according to the image characteristics and the task type includes:

Performing code conversion on the image features and the task types to obtain corresponding feature codes and task type codes, and calculating Euclidean distances between the feature codes and the task type codes and model performance codes of all pre-labeling models in a preset model library;

And determining the matching degree of the Euclidean distance and each pre-labeling model, and determining the model with the largest matching degree as the corresponding pre-labeling model.

Further, the storing the image data after the pre-labeling processing according to the data type by the corresponding standardized label includes:

Storing image data pre-marked as target detection, key point detection and instance segmentation labels in a Coco standard format;

Image data pre-labeled as image classification labels are stored in OpenMMLab standard format.

Further, the converting the image processing model into a corresponding executable format according to the specific type of the target deployment hardware and converting the floating point parameter number of the image processing model into an integer or fixed point number, to obtain an image processing model after format conversion and parameter quantization, including:

Adjusting the arrangement mode of data in the image processing model according to the specific type of the target deployment hardware;

And converting the point value into an integer value according to the quantization factor and the offset of the image processing model to obtain the image processing model after format conversion and parameter quantization.

Further, the inputting, by the inference interface, the image data stored in the standardized tag after being processed by the preprocessing interface to the image processing model, and processing, by the post-processing interface, a model output of the image processing model, and determining, according to the model output, a corresponding model precision, includes:

Processing the image data stored by the standardized tag through the preprocessing interface, inputting the image data processed by the preprocessing interface into the image processing model through the reasoning interface, and operating the image processing model;

And processing the model output of the target deployment hardware after the image processing model is operated through the post-processing interface, and determining the corresponding model precision according to the model output.

In a second aspect, the present application provides a one-stop artificial intelligence image processing model construction apparatus, comprising:

The pre-labeling module is used for creating an image labeling task, determining a corresponding pre-labeling model according to the image labeling task, pre-labeling preset image data according to the pre-labeling model, and storing the image data subjected to the pre-labeling according to the data type;

The training evaluation module is used for receiving target deployment hardware and target model running speeds selected by a user, determining the structure of a corresponding image processing model according to the numerical comparison relation between the pre-evaluation speed of each preset image processing model on the target deployment hardware and the target model running speed, and training the image processing model according to the image data stored by the standardized label to obtain the image processing model;

The quantization conversion module is used for converting the image processing model into a corresponding executable format according to the specific type of the target deployment hardware and converting floating point parameters of the image processing model into integers or fixed point numbers to obtain the image processing model after format conversion and parameter quantization;

The test deployment module is used for determining a pre-processing interface, an inference interface and a post-processing interface corresponding to the target deployment hardware according to a model definition specification corresponding to the image processing model, inputting the image data stored by the standardized tag processed by the pre-processing interface to the image processing model through the inference interface, processing the model output of the image processing model through the post-processing interface, determining corresponding model precision according to the model output, and deploying the image processing model to the target deployment hardware when the model precision accords with a preset precision condition.

In a third aspect, the present application provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the steps of the one-stop artificial intelligence image processing model building method when executing the program.

In a fourth aspect, the present application provides a computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps of the one-stop artificial intelligence image processing model building method.

In a fifth aspect, the present application provides a computer program product comprising computer programs/instructions which when executed by a processor implement the steps of the one-stop artificial intelligence image processing model building method.

According to the technical scheme, the application provides a one-stop artificial intelligent image processing model construction method and device, an image labeling task is established, the structure of a corresponding image processing model is determined, and training of the image processing model is carried out according to image data stored by a standardized label, so that the image processing model is obtained; according to the model definition specification corresponding to the image processing model, a corresponding preprocessing interface, an inference interface and a post-processing interface of the target deployment hardware are determined, image data stored by a standardized label processed by the preprocessing interface is input to the image processing model through the inference interface, model output of the image processing model is processed through the post-processing interface, corresponding model precision is determined according to the model output, and the image processing model is deployed to the target deployment hardware when the model precision meets the preset precision condition, so that the integrated integration of full flows of data labeling, model training, model conversion, model deployment and the like can be realized, development efficiency is improved, development cost is reduced, and high consistency of training precision and deployment precision is ensured.

Drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described, and it is obvious that the drawings in the following description are some embodiments of the present application, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic flow diagram of a one-stop artificial intelligence image processing model construction method according to an embodiment of the application;

FIG. 2 is a second flow chart of a method for constructing a one-stop artificial intelligence image processing model according to an embodiment of the application;

FIG. 3 is a third flow chart of a method for constructing a one-stop artificial intelligence image processing model according to an embodiment of the application;

FIG. 4 is a flow chart of a method for constructing a one-stop artificial intelligence image processing model according to an embodiment of the present application;

FIG. 5 is a flowchart of a method for constructing a one-stop artificial intelligence image processing model according to an embodiment of the present application;

FIG. 6 is a flowchart of a method for constructing a one-stop artificial intelligence image processing model according to an embodiment of the present application;

FIG. 7 is a flow chart of a method for constructing a one-stop artificial intelligence image processing model according to an embodiment of the application;

FIG. 8 is a block diagram of a one-stop artificial intelligence image processing model building apparatus in an embodiment of the application;

Fig. 9 is a schematic structural diagram of an electronic device in an embodiment of the application.

Reference numerals:

an electronic device 9600, a central processor 9100, a memory 9140, a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, a power supply 9170, a buffer memory 9141, an application/function storage portion 9142, a data storage portion 9143, a driver storage portion 9144, an antenna 9111, a speaker 9131, and a microphone 9132.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the embodiments of the present application more apparent, the technical solutions of the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application, and it is apparent that the described embodiments are some embodiments of the present application, but not all embodiments of the present application. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.

The technical scheme of the application obtains, stores, uses, processes and the like the data, which all meet the relevant regulations of national laws and regulations.

The problems that in the prior art, the artificial intelligent model development process has low data labeling efficiency, low model training efficiency, complex model conversion, unaligned deployment precision and the like are solved. The application provides a one-stop artificial intelligent image processing model construction method and device, which not only increase the complexity and cost of artificial intelligent model development, but also limit the popularization and promotion of artificial intelligent technology in practical application, and the structure of a corresponding image processing model is determined by creating an image labeling task, and training of the image processing model is carried out according to image data stored by a standardized label to obtain the image processing model; according to the model definition specification corresponding to the image processing model, a corresponding preprocessing interface, an inference interface and a post-processing interface of the target deployment hardware are determined, image data stored by a standardized label processed by the preprocessing interface is input to the image processing model through the inference interface, model output of the image processing model is processed through the post-processing interface, corresponding model precision is determined according to the model output, and the image processing model is deployed to the target deployment hardware when the model precision meets the preset precision condition, so that the integrated integration of full flows of data labeling, model training, model conversion, model deployment and the like can be realized, development efficiency is improved, development cost is reduced, and high consistency of training precision and deployment precision is ensured.

In order to realize the integrated integration of the whole processes of data labeling, model training, model conversion, model deployment and the like, improve the development efficiency, reduce the development cost and ensure the high consistency of the training precision and the deployment precision, the application provides an embodiment of a one-stop artificial intelligent image processing model construction method, referring to fig. 1, wherein the one-stop artificial intelligent image processing model construction method specifically comprises the following contents:

step S101: creating an image annotation task, determining a corresponding pre-annotation model according to the image annotation task, performing pre-annotation processing on preset image data according to the pre-annotation model, and storing the image data subjected to the pre-annotation processing in a corresponding standardized tag according to a data type;

In the embodiment of the application, an automatic image labeling process is described in detail, and aims to improve the processing efficiency and labeling accuracy of image data through a pre-labeling model. First, the creation of the image annotation task is the starting point of the overall process. In particular, the present embodiment needs to specify specific requirements of image labeling, such as identifying objects in an image, performing scene classification, performing image segmentation, or the like. The task definition at this stage will directly affect the selection and training strategy of the subsequent pre-labeling model.

After the task definition is clear, the next step in this embodiment is to collect image data suitable for the preset image annotation task. These image data may originate from existing data sets or be obtained by new data acquisition procedures, it being critical to ensure that the data can cover all relevant variables and scenes so that the model can learn how to make efficient image labeling in various situations.

Then, according to the requirements of the image labeling task, the embodiment selects or develops a suitable pre-labeling model. For example, if the task is object detection, a convolutional neural network model in deep learning, such as YOLO or SSD, may be selected for use; for image segmentation tasks, models such as U-Net or Mask R-CNN may be more suitable. After the model is selected, the embodiment trains the model, and the labeled data set is used for training the model so as to ensure that the model can accurately predict and label the unseen new image.

After model training is completed, the embodiment applies the pre-labeling model to the collected preset image data. In this way, the model automatically identifies and annotates relevant features or objects in the image, thereby generating preliminary tag data. The method not only improves the labeling efficiency, but also has consistency and repeatability compared with manual labeling because the labeling is performed by a machine learning model.

Next, the present embodiment performs normalization processing on the tag data generated by the model to ensure consistency of the data format and the tag. The standardized process includes unifying the naming, format and storage structure of the tags, which is critical to subsequent data processing and analysis. After the normalization process is completed, the present embodiment stores the annotation data in a database or in a specific file system. The storage mode is generally dependent on the requirements of the subsequent application, and may be selected in a structured database format or unstructured file storage mode.

In general, the method and the device for marking the images by using the pre-marking model to automate the image marking process remarkably improve marking efficiency and accuracy, reduce errors caused by human factors, improve data consistency and standardization level, and provide high-quality data support for subsequent machine learning training and image analysis tasks. In addition, the method of the embodiment is easy to expand to a larger data set and a more complex labeling task, is suitable for large-scale image data processing and analysis, and has wide application prospects.

Step S102: receiving target deployment hardware and target model running speeds selected by a user, determining the structure of a corresponding image processing model according to the numerical comparison relation between the pre-evaluation speed of each preset image processing model on the target deployment hardware and the target model running speed, and training the image processing model according to image data stored by the standardized label to obtain the image processing model;

In an embodiment of the present application, step S102 involves receiving a user selected target deployment hardware and target model operating speed. The core purpose of this step is to ensure that the selected image processing model not only meets the functional requirements, but also achieves the desired performance on the hardware platform specified by the user. By considering the computing capacity of the target deployment hardware and the running speed of the target model, the image processing model which is most suitable for the user requirement can be effectively screened out.

Specifically, the present embodiment first requires collection of user detailed information about the target deployment hardware, including but not limited to processor type, GPU acceleration capability, memory size, and other related hardware configurations. This information is critical to evaluating the performance of different image processing models on this hardware. In addition, the user also needs to provide specific requirements for the speed of operation of the target model, such as the number of frames per second of processed image or specific requirements for delay.

After these basic information are acquired, the present embodiment will evaluate the running performance of a plurality of image processing models set in advance on the target deployment hardware. This evaluation process typically depends on the complexity of the model, the computational resources required, and the computational power of the hardware. The evaluation may be done by actually running the models or by simulation software in order to get the predicted speed of each model under a specific hardware configuration.

Then, the embodiment determines the most suitable image processing model according to the numerical comparison relation between the running speed of the target model and the pre-evaluation speed of each preset image processing model on the target deployment hardware. This comparison is based on whether the user-set speed requirement is met. For example, if the user requires high speed processing, then the model that runs the fastest on a given hardware is selected; conversely, if the speed requirements are not very stringent, a model with a better balance between performance and processing speed may be selected.

The implementation of the step ensures that the image processing system can reach expected performance in actual deployment, optimizes the use of resources, ensures that the system meets performance requirements and simultaneously keeps reasonable cost and energy efficiency. The finally determined image processing model provides powerful support for subsequent image processing tasks, and ensures that the system can work efficiently and stably in actual operation. In this way, the present embodiment effectively combines user requirements with technical capabilities, providing a highly customized solution to meet specific requirements of different users in diverse application scenarios.

Step S103: converting the image processing model into a corresponding executable format according to the specific type of the target deployment hardware, and quantizing floating point parameters of the image processing model into integers or fixed point numbers to obtain an image processing model subjected to format conversion and parameter quantization;

In step S103 in the embodiment of the present application, it is described that the selected image processing model is converted into an executable format suitable for the target deployment hardware, and quantization processing of parameters is performed to optimize the execution efficiency and hardware compatibility of the model. This step is a critical loop in the model deployment process that ensures that the model can run efficiently on specific hardware while reducing resource consumption.

First, the present embodiment needs to consider the specific type of target deployment hardware, including but not limited to CPU, GPU, FPGA or dedicated neural network accelerators, etc. Different types of hardware have different requirements on the executable format of the model. For example, GPUs typically require conversion of a model into a format compatible with CUDA or OpenCL, whereas FPGAs may require conversion of a model into a specific Hardware Description Language (HDL) format.

Next, the present embodiment will perform format conversion of the model. This process involves converting the model from its original developed format (e.g., tensorFlow or PyTorch format) to a format supported by the target hardware. This is typically accomplished by using specialized conversion tools, such as TensorFlow Lite Converter, ONNX converters, and the like. These tools are able to read the original model and output a model format compatible with the target hardware.

In addition to format conversion, step S103 also includes quantizing the floating point parameters in the model to integers or fixed point numbers. Quantization is a commonly used model optimization technique, and reduces the storage and calculation requirements of a model by reducing the data precision of model parameters, thereby increasing the running speed and reducing the power consumption. In this process, floating point parameters are converted to integer or fixed point numbers, which typically involve some degree of loss of precision, but can significantly improve the efficiency of model execution on resource-constrained hardware.

And finally, the obtained image processing model subjected to format conversion and parameter quantization can be deployed on target hardware for execution. The optimized model not only can meet the performance requirement, but also can exert the computing capacity of hardware to the greatest extent while ensuring the operation efficiency.

By implementing step S103, the present embodiment effectively converts the high-level image processing model into an optimized execution version customized for a specific hardware environment, ensuring the practicality and high efficiency of the model. The method is not only beneficial to reducing the deployment cost and the running cost, but also enables the image processing system to stably run in various different hardware environments, and meets wider application requirements.

Step S104: determining a pre-processing interface, an inference interface and a post-processing interface corresponding to the target deployment hardware according to a model definition specification corresponding to the image processing model, inputting image data stored by the standardized tag after being processed by the pre-processing interface to the image processing model through the inference interface, processing model output of the image processing model through the post-processing interface, determining corresponding model precision according to the model output, and deploying the image processing model to the target deployment hardware when the model precision meets a preset precision condition.

In step S104, the embodiment of the present application details how to integrate the image processing model with the target deployment hardware, so as to ensure that the model can reach the preset performance and precision requirements in the actual operating environment. This process involves the explicit definition of the interface, the processing of the input-output data, and the final deployment and accuracy verification of the model.

Firstly, according to the definition specification of the image processing model, the embodiment of the application needs to define the interface types of the implementation needed by the target deployment hardware, including a preprocessing interface, an inference interface and a post-processing interface. The preprocessing interface is mainly responsible for performing necessary preprocessing operations on the input image data, such as format conversion, normalization, clipping and the like, so as to ensure that the data format is matched with the input requirements of the model. The inference interface is then connected to the core computing resources of the hardware, such as a CPU, GPU or other dedicated accelerator, responsible for executing the computational tasks of the model. The post-processing interface is used for processing the output data of the model, executing tasks such as decoding, non-maximum suppression and the like, and converting the original output into a result format which is available in practical application.

Next, the embodiment of the present application inputs the standardized image data processed through the preprocessing interface into the image processing model. This process typically requires handling data buffering and transmission issues to ensure that the data is efficiently and accurately delivered to the model for analysis. After the model is processed, the output data is sent to a post-processing interface. In the post-processing interface, the output data is converted into specific application results, for example, in an object detection model, the post-processing step screens out detection frames with high confidence and calculates the positions and categories of the frames.

After finishing the post-processing, embodiments of the present application will determine the accuracy of the model based on the model output. This step is performed by comparing the model output with the expected result or standard answer to evaluate whether the model meets a preset accuracy criterion. If the precision of the model meets or exceeds the preset condition, the embodiment of the application can formally deploy the image processing model to the target deployment hardware.

This deployment process includes not only physically configuring the model code and parameters onto the hardware, but also ensuring that all relevant systems and interfaces are ready to support continued operation and maintenance of the model. In addition, after deployment, field tests are required to verify whether the performance of the model in the real environment meets the expectations or not, so that the stability and the reliability of the system are ensured.

Through the steps, the embodiment of the application ensures that the image processing model not only has high precision in theory, but also can be efficiently and accurately executed in practical application, thereby meeting the requirements of various application scenes. This systematic approach promotes the practicality of the model and the feasibility of industrial applications, which are key steps in the commercialization and practicality of modern image processing technology.

As can be seen from the above description, the one-stop artificial intelligent image processing model construction method provided by the embodiment of the application can determine the structure of the corresponding image processing model by creating an image labeling task, and perform training of the image processing model according to the image data stored in the standardized label to obtain the image processing model; according to the model definition specification corresponding to the image processing model, a corresponding preprocessing interface, an inference interface and a post-processing interface of the target deployment hardware are determined, image data stored by a standardized label processed by the preprocessing interface is input to the image processing model through the inference interface, model output of the image processing model is processed through the post-processing interface, corresponding model precision is determined according to the model output, and the image processing model is deployed to the target deployment hardware when the model precision meets the preset precision condition, so that the integrated integration of full flows of data labeling, model training, model conversion, model deployment and the like can be realized, development efficiency is improved, development cost is reduced, and high consistency of training precision and deployment precision is ensured.

In an embodiment of the one-stop artificial intelligence image processing model construction method of the present application, referring to fig. 2, the following may be further specifically included:

step S201: determining a corresponding pre-labeling model according to the task type of the image labeling task and the data characteristics of preset image data;

Step S202: and inputting the preset image data into the pre-labeling model for pre-labeling treatment to obtain pre-labeled image data.

In step S201, the embodiment of the present application first needs to determine an appropriate pre-labeling model according to a specific type of an image labeling task and a feature of preset image data. Image annotation tasks may include, but are not limited to, object detection, image classification, semantic segmentation, and the like. The requirements for the models are different for each task type, for example, the object detection task may need to select a model that can identify and locate multiple objects in the image, while the image classification task may need a model that can properly classify the entire image. The selection of the pre-annotation model is affected by the characteristics of the pre-set image data, such as the image size, resolution, color depth, and complexity of the objects in the image. Therefore, when the pre-labeling model is determined, the factors are comprehensively considered, and the model which is most suitable for the current task type and the data characteristic is selected so as to ensure the accuracy and the efficiency of the pre-labeling.

In step S202, the embodiment of the present application inputs preset image data into the determined pre-labeling model for processing. In this process, the preset image data is first subjected to necessary preprocessing, such as adjusting the image size and normalizing the image pixel values, so as to meet the input requirement of the pre-labeling model. The preprocessed image data is sent to a pre-labeling model, which automatically analyzes the image content and generates a preliminary labeling result. These labeling results may include bounding boxes of objects, class labels, object masks at the pixel level, etc., with the specific content being dependent on the type of image labeling task.

The obtained pre-labeling image data can provide a basis for the follow-up detailed labeling work, and the workload and time cost of manual labeling are greatly reduced. For example, in the process of creating a training data set of an automatic driving system, a pre-labeling model can automatically identify and label key objects such as vehicles, pedestrians and the like in images, and then labeling personnel only need to correct and refine the automatically generated labels, and each image does not need to be labeled from scratch. The pre-labeling processing not only improves the labeling efficiency, but also improves the consistency and quality of the data set.

Through these two steps, embodiments of the present application demonstrate how to efficiently use pre-labeling models to process and prepare image data for more in-depth analysis or machine learning training. The method is particularly suitable for scenes needing to process a large amount of image data, such as machine learning projects, large-scale image analysis tasks and the like, and can remarkably improve the efficiency and quality of data preparation.

In an embodiment of the one-stop artificial intelligence image processing model construction method of the present application, referring to fig. 3, the following may be further specifically included:

Step S301: determining task types of the image labeling task, and determining image recognition emphasis features according to the task types, wherein the task types comprise an image classification task, an image detection task and an image segmentation task, and the image recognition emphasis features comprise emphasis global features and emphasis local features;

step S302: and extracting key features of preset image data to obtain corresponding image features, and matching corresponding pre-labeling models from a preset model library according to the image features and the task types.

In step S301, the embodiment of the present application accurately identifies and classifies the type of the image labeling task. The types of labeling tasks mainly comprise three types: an image classification task, an image detection task, and an image segmentation task. Each task type corresponds to a different emphasis point and processing requirement, and therefore, for each task type, the present embodiment will determine the feature to which image recognition should be emphasized according to its characteristics.

Specifically, the image classification task focuses mainly on the global features of the image. Global features include the overall composition, color distribution, texture, shape, etc. of the image, which features help identify the scene or object class represented by the image. For example, in a classification task for animal identification, the model needs to capture global information, such as body type, color, and typical pose, that is sufficient to distinguish different animal species.

The image detection task is then focused on local features, as it requires not only identifying objects in the image, but also determining the specific location and extent of these objects. Local features relate to edges, corners, texture details, etc. of objects, which are critical for accurately locating and identifying individual objects in an image. For example, in vehicle detection, a detection model needs to be able to identify key local information such as the contour of a vehicle, windows, and license plates.

The emphasis of the image segmentation task is also a local feature, but unlike the detection task, the segmentation task needs to operate on a pixel level, classifying each pixel in the image into a corresponding object or background class. This requires that the model be able to understand and process detailed information such as the fine boundaries, textures, and spatial relationships of objects with other objects.

In step S302, the embodiment of the present application performs detailed key feature extraction on the preset image data. This step is a critical loop in the overall image processing flow, as the extracted features will directly affect the quality of matching and pre-labeling of the subsequent models. Feature extraction methods typically involve the use of advanced image processing algorithms, such as edge detection, feature point extraction, texture analysis, etc., to accurately capture features that are focused on the corresponding task type.

For example, when processing an image classification task, feature extraction may use global descriptors, such as color histograms, GIST descriptors, etc., that can capture the overall visual content of the image. In the image detection or segmentation task, local feature extraction techniques, such as SIFT, HOG or convolution feature map in deep learning, may be used, which can describe the appearance and shape of each local region in the image in detail.

After feature extraction is completed, the embodiment of the application matches a proper pre-labeling model from a preset model library according to the extracted image features and the determined task types. The model library contains a plurality of models optimized for different image processing tasks, each model being trained and tailored to specific feature combinations and task requirements. By comparing the image characteristics with the characteristics of each model in the model library, the embodiment can select the model which is most suitable for the current image data and the labeling requirements.

After selecting a correct pre-labeling model, the embodiment of the application inputs preset image data into the model to perform automatic pre-labeling processing. The process not only greatly improves the labeling efficiency, but also ensures the consistency and accuracy of labeling results. The application of the pre-labeling model particularly in a large-scale image data processing scene obviously reduces the requirement of manual labeling and the related cost, and improves the automation level of data processing.

In addition, when the pre-labeling model is selected, the generalization capability and the processing speed of the model are also considered, so that the model is ensured to be excellent in a specific data set, and the model can adapt to changeable actual application scenes and data environments. This is particularly important for image processing systems that are used in a practical production environment, because image data in practical applications tends to be more complex and variable than standard data sets.

In actual practice, step S302 of the embodiment of the present application is not limited to pre-labeling using a single model. In some cases, in order to further improve accuracy and robustness of pre-labeling, the embodiment may use a model fusion technique, and combine prediction results of multiple models to make a final labeling decision. The method can effectively reduce possible deviation and deficiency of a single model and improve the overall prediction performance.

Model fusion typically employs a variety of techniques including, but not limited to, voting systems, average prediction, weighted averaging, and more complex ensemble learning methods such as Boosting and Bagging. Through the strategies, the expertise of different models on specific tasks can be synthesized, and more accurate and reliable pre-marking can be carried out on the image data.

In addition, in the process of model matching and selecting, the embodiment of the application also considers updating and iteration of the model. Over time and accumulation of data, the original model may need to be updated to accommodate new data features or to improve algorithms. Therefore, the embodiment is provided with a model updating mechanism, the performance of the model is periodically evaluated, and the model library is maintained and updated according to actual requirements and technical progress.

In the whole image labeling process, the embodiment of the application realizes high-efficiency and high-quality automatic image labeling through accurate task type judgment, detailed feature extraction and intelligent model matching and fusion. The process not only optimizes the use of resources and reduces the labor cost, but also provides a reliable data base for subsequent image analysis and application.

In summary, the embodiment of the application provides an image pre-labeling method based on task type and feature emphasis, which greatly improves the efficiency and quality of image labeling through accurate matching and intelligent processing. The method has wide application prospect in various image processing fields such as automatic driving, medical image analysis, security monitoring and the like, and can effectively support the rapid development and technical innovation of the fields.

In an embodiment of the one-stop artificial intelligence image processing model construction method of the present application, referring to fig. 4, the following may be further specifically included:

Step S401: performing code conversion on the image features and the task types to obtain corresponding feature codes and task type codes, and calculating Euclidean distances between the feature codes and the task type codes and model performance codes of all pre-labeling models in a preset model library;

Step S402: and determining the matching degree of the Euclidean distance and each pre-labeling model, and determining the model with the largest matching degree as the corresponding pre-labeling model.

In step S401, the embodiment of the present application first performs transcoding of image features and task types. The purpose of this process is to convert the visual features of the image and the class of annotation tasks into a mathematically processable format, i.e. feature encoding and task type encoding. This transcoding is critical to the subsequent model matching process, as it ensures that different types of data can be compared and analyzed under the same mathematical framework.

Feature encoding typically involves converting visual features (e.g., color, texture, shape, etc.) of an image into points in a vector space. This may be achieved by various feature extraction algorithms, for example using the middle layer output of the deep learning model as a feature descriptor. Task type coding involves converting the class of tasks (e.g., classification, detection, or segmentation) into a standard form, typically a single-hot coded vector, where each task type is fixed in position in the vector with corresponding positions of 1 and the remainder of 0.

After the coding is completed, the Euclidean distance between the calculated feature codes and task type codes and the model performance codes of all pre-marked models in a preset model library is exemplified. Model performance coding is the result of coding the performance characteristics of each pre-labeled model, reflecting the efficiency and accuracy of the model in handling a particular type of task. The euclidean distance is calculated as:

d (p, q) = Σi=1n (qi-pi) 2 where p and q are vectors of feature encoding (combined task type encoding) and model performance encoding, respectively, and n is the dimension of the vector.

By calculating Euclidean distance, the present examples are able to evaluate the similarity of each pre-labeled model to the current image and task type. The smaller the distance, the smaller the difference between the codes, and the higher the matching degree of the model and the task.

In step S402, according to the calculated euclidean distance, the embodiment of the present application determines the matching degree of each pre-labeling model. The degree of matching can generally be quantified by the inverse of the distance or other function, with higher degrees of matching indicating better suitability of the model for the current task.

In this step, the embodiment of the application selects the model with the largest matching degree as the final pre-labeling model. This means that of all available pre-labeling models, the model closest to the current image feature and task type is selected to perform the pre-labeling task. The selection method ensures the high efficiency of the labeling process and the accuracy of the labeling result.

After the optimal pre-labeling model is selected, the model is used for automatically labeling the input image data, and a preliminary labeling result is generated. This result may be used directly for subsequent image processing tasks or as a basis for manual review and fine tuning.

In general, steps S401 and S402 constitute an efficient model matching and selection process that not only optimizes the selection and application of the model, but also ensures the performance and reliability of the overall image annotation system. By the method, the embodiment of the application can realize automatic, high-efficiency and high-accuracy labeling in various image labeling tasks, and support complex image processing requirements and applications.

In an embodiment of the one-stop artificial intelligence image processing model construction method of the present application, referring to fig. 5, the following may be further specifically included:

Step S501: storing image data pre-marked as target detection, key point detection and instance segmentation labels in a Coco standard format;

Step S502: image data pre-labeled as image classification labels are stored in OpenMMLab standard format.

In step S501, image data relating to object detection, keypoint detection, and instance segmentation is stored using COCO (Common Objects in Context) standard format. The COCO format is a data format widely used in the field of computer vision and is well suited for storing images and associated complex annotation information (e.g., bounding boxes, segmentation masks, etc.). The COCO format generally includes the following components:

images contains basic information of the image, such as file name, image size, etc.

Annotations detailed labeling information for objects in the image, such as bounding boxes for object detection, locations of keypoints, polygons for instance segmentation, etc.

And (3) listing all object categories and IDs thereof for consistency of classification and labeling.

This format supports the integration of multiple types of annotation data in a unified file (typically JSON format), facilitating processing and sharing.

Step S502 is to specially process pre-labeled data of the image classification task, and store the pre-labeled data in OpenMMLab standard format. OpenMMLab is an open-source computer vision algorithm research platform that provides a variety of deep learning resources and tools. For image classification tasks, classification tags and image files can be efficiently integrated and managed using OpenMMLab formats, which typically include:

Image path: pointing to the path where the image is stored.

And (3) tag: a category label associated with each image.

The OpenMMLab format typically employs a compact structure, such as a text file or JSON file, in which the paths of the image files and corresponding tags are listed, which facilitates rapid loading and processing of large-scale image classification datasets.

By adopting the two standardized data formats, the consistency and compatibility of the data can be ensured, and meanwhile, various open source tools and libraries can be conveniently used for further data processing and model training. The method improves the usability and flexibility of the pre-labeling data and lays a solid foundation for the subsequent image analysis task.

In an embodiment of the one-stop artificial intelligence image processing model construction method of the present application, referring to fig. 6, the following may be further specifically included:

step S601: adjusting the arrangement mode of data in the image processing model according to the specific type of the target deployment hardware;

step S602: and converting the point value into an integer value according to the quantization factor and the offset of the image processing model to obtain the image processing model after format conversion and parameter quantization.

In step S601, it is important to adjust the arrangement of data in the image processing model according to the specific characteristics of the target deployment hardware. The purpose of this step is to ensure that the layout of the data in memory is best suited to the hardware architecture used, thereby achieving maximum efficiency and performance in performing the image processing tasks. Different hardware platforms, such as CPU, GPU, FPGA or ASICs, each have their own specific data access optimization strategy. For example, the GPU may be more optimized to process NHWC (lot, height, width, channel) formatted data because the GPU is able to access consecutive memory locations faster when processing parallel operations. While CPUs may prefer to use NCHW (lot, channel, height, width) formats because such data placement may reduce cache miss situations, thereby improving efficiency.

The advantage of converting the data format to accommodate particular hardware is not merely reflected in speed. It also significantly reduces power consumption because the optimized memory access pattern reduces the number of data moves that must be performed, each of which consumes power. Furthermore, proper data placement can improve program scalability, making it easier and efficient to extend or migrate to other hardware in the future.

The implementation of step S602 is to convert floating point values in the image processing model into integer values by quantization techniques. Quantization is a technique commonly used for deep learning model optimization, especially when the model is deployed on edge devices. Edge devices typically have limited computational resources and memory space, while quantization can significantly reduce model size and computational requirements. In quantization, each floating point value is converted to an integer by multiplying by a quantization factor and adding an offset. The quantization factor determines the degree of scaling of the values, and the offset is used to adjust the range of values after conversion.

Quantization not only reduces the memory size of the model, but also improves the efficiency of operation, as integer operations are typically faster than floating point operations. Furthermore, integer arithmetic may also reduce power consumption, which is particularly important for applications running on battery powered devices. However, quantization may also involve a loss of accuracy, so selecting an appropriate quantization strategy and adjusting quantization parameters is critical to ensure that the model still maintains good performance.

In performing these steps, a number of factors need to be considered. First, the developer needs to know the architecture and optimization strategy of the target hardware in detail. This typically involves reading a document of the hardware manufacturer and possibly a hardware specification. Second, the quantization strategy should be selected based on the specific use of the model and the required accuracy criteria. In some high precision demanding applications it may be desirable to employ finer quantization strategies or to use mixed precision methods at some level.

In the whole process, sufficient tests should also be performed to ensure that the adjusted model can achieve the expected performance in actual deployment. This includes running the model on the target hardware, monitoring its execution speed, accuracy, and resource consumption, among others. Furthermore, simulation tools may also be used to predict the behavior of hardware in actual operation in order to discover and solve potential problems before the model is fully deployed.

Through these two steps, adjusting the data arrangement and model quantification, plays a key role in the deployment of deep learning models, especially where the target hardware has resource limitations. The steps not only can optimize the performance of the model, but also can greatly improve the feasibility and the efficiency of the model in practical application.

In an embodiment of the one-stop artificial intelligence image processing model construction method of the present application, referring to fig. 7, the following may be further specifically included:

Step S701: processing the image data stored by the standardized tag through the preprocessing interface, inputting the image data processed by the preprocessing interface into the image processing model through the reasoning interface, and operating the image processing model;

Step S702: and processing the model output of the target deployment hardware after the image processing model is operated through the post-processing interface, and determining the corresponding model precision according to the model output.

The preprocessing interface involved in step S701 is typically designed to ensure that the input data meets the requirements of the image processing model. In this step, it is first necessary to process the image data stored in the standardized tag. This typically involves several key operations: resizing, format conversion, normalization or normalization of the image and possibly data enhancement techniques. For resizing, it is often necessary to resize the input image to the size used in model training, as deep learning models have stringent requirements on the input size. Format conversion involves converting an image file from its original format to a format that the model can handle, such as converting JPEG to PNG or converting a non-standard color space to RGB. Normalization or normalization is the scaling of the pixel values of the image data to a specific range of values, such as 0 to 1 or-1 to 1, which helps the model to learn better and converge quickly.

Data enhancement is an optional step that is typically used in the training phase to increase the generalization ability of the model, but can also be used in the reasoning phase to increase the robustness of the model. The data enhancement may include techniques of rotation, scaling, cropping, color adjustment, etc. These techniques not only simulate different operating conditions, but also increase the diversity of data, thereby enabling the model to better handle real world situations.

After the preprocessing is completed, the processed image data is input into an image processing model through an inference interface. The inference interface is responsible for formatting the pre-processed data into a form acceptable to the model and ensuring the correctness and efficiency of the data in the transmission process. Then, an image processing model is run, which analyzes and processes the input data according to the learned features, and outputs a prediction result.

Step S702 focuses mainly on the determination of the post-processing interface and the model accuracy. Once the image processing model has completed reasoning, its output typically requires further post-processing to be converted into useful information or visual results. Post-processing tasks may include decoding the output of the model (e.g., converting the probabilistic output of the classification task into class labels), adapting the output format to the needs of the downstream system, or making a visual presentation of the results.

In addition, post-processing of the model output may involve correction or optimization of the results, such as adjusting the original output of the model using certain rules or heuristics to improve its actual effectiveness in a particular application. For example, in a face recognition system, a threshold may be introduced to decide when to accept or reject recognition results, or in traffic sign recognition, additional logic may be performed to identify multiple candidate signs for a model.

Finally, determining the accuracy of the model is an important step in evaluating the performance of the model. Typically, this involves comparing the model output with a set of pre-labeled real data to calculate various performance metrics such as accuracy, recall, F1 score, etc. The evaluation of the model precision can help a developer and a user to know the performance of the model in practical application, and can provide guidance for further optimization of the model.

In the whole process, it is important to ensure that data streams are efficiently and safely transmitted between different interfaces. Furthermore, the implementation of the various process steps requires careful design to ensure that the scalability and maintainability of the process is maintained while meeting performance requirements. The specific implementation of each step may vary according to different application scenarios and performance characteristics of the target hardware.

For example, in designing pre-processing and post-processing interfaces, the computational power and memory limitations of the target deployment hardware need to be considered. On some resource constrained devices, optimization algorithms may be required to reduce computational effort and memory usage, such as by simplifying data enhancement steps or using lighter weight image processing algorithms. In addition, for certain applications, such as real-time video surveillance or image recognition on mobile devices, there is a need to optimize the data processing flow to reduce latency and increase response speed.

In model reasoning, it is also important to ensure that the model is able to run efficiently. This typically involves techniques such as quantization, pruning and compression of the model, with the aim of reducing the size of the model and increasing the speed of operation, while maintaining the accuracy of the model as much as possible. For deep learning models, specialized hardware accelerators, such as GPUs or TPUs, may also be utilized to increase processing speed.

For the post-processing step, in addition to basic output parsing and format conversion, more complex image post-processing techniques may be included, such as image denoising, sharpening, super resolution, etc., which may further improve the quality of the output image to make it more suitable for specific application requirements.

The precision assessment is an important link after model deployment. This involves not only comparison with standard data sets, but also continuous performance monitoring in the actual application environment. In certain applications, such as medical image analysis or safety monitoring systems, reliability of the model is particularly important. Thus, in addition to the initial accuracy assessment, periodic re-assessment and calibration of the model is required to accommodate possible environmental changes or data drift.

In general, steps S701 and S702 constitute a complete flow from preprocessing of image data to model reasoning to post-processing and accuracy assessment. This process not only requires a scientific and efficient technical implementation, but also takes into account the characteristics and limitations of the actual deployment environment. By careful design and optimization of these steps, it can be ensured that the image processing model meets the established performance goals and application requirements in practical applications, effectively serving the end user.

In order to achieve the integration of the whole processes of data labeling, model training, model conversion, model deployment and the like, improve the development efficiency, reduce the development cost and ensure the high consistency of the training precision and the deployment precision, the application provides an embodiment of a one-stop artificial intelligent image processing model construction device for realizing all or part of the content of the one-stop artificial intelligent image processing model construction method, and referring to fig. 8, the one-stop artificial intelligent image processing model construction device specifically comprises the following contents:

the pre-labeling module 10 is used for creating an image labeling task, determining a corresponding pre-labeling model according to the image labeling task, performing pre-labeling processing on preset image data according to the pre-labeling model, and storing the image data subjected to the pre-labeling processing in a corresponding standardized label according to a data type;

The training evaluation module 20 is configured to receive the target deployment hardware and the target model running speed selected by the user, determine a structure of a corresponding image processing model according to a numerical comparison relation between a pre-evaluation speed of each preset image processing model on the target deployment hardware and the target model running speed, and perform training of the image processing model according to the image data stored in the standardized label, so as to obtain the image processing model;

The quantization conversion module 30 is configured to convert the image processing model into a corresponding executable format according to the specific type of the target deployment hardware, and quantize the floating point parameter of the image processing model into an integer or a fixed point number, so as to obtain an image processing model after format conversion and parameter quantization;

The test deployment module 40 is configured to determine a pre-processing interface, an inference interface and a post-processing interface corresponding to the target deployment hardware according to a model definition specification corresponding to the image processing model, input, through the inference interface, image data stored by the standardized tag after being processed by the pre-processing interface to the image processing model, process, through the post-processing interface, a model output of the image processing model, determine a corresponding model precision according to the model output, and deploy the image processing model to the target deployment hardware when the model precision meets a preset precision condition.

As can be seen from the above description, the one-stop artificial intelligent image processing model construction device provided by the embodiment of the application can determine the structure of the corresponding image processing model by creating an image labeling task, and perform training of the image processing model according to the image data stored in the standardized label to obtain the image processing model; according to the model definition specification corresponding to the image processing model, a corresponding preprocessing interface, an inference interface and a post-processing interface of the target deployment hardware are determined, image data stored by a standardized label processed by the preprocessing interface is input to the image processing model through the inference interface, model output of the image processing model is processed through the post-processing interface, corresponding model precision is determined according to the model output, and the image processing model is deployed to the target deployment hardware when the model precision meets the preset precision condition, so that the integrated integration of full flows of data labeling, model training, model conversion, model deployment and the like can be realized, development efficiency is improved, development cost is reduced, and high consistency of training precision and deployment precision is ensured.

In order to realize the integrated integration of the whole processes of data labeling, model training, model conversion, model deployment and the like in terms of hardware, improve development efficiency, reduce development cost and ensure high consistency of training precision and deployment precision, the application provides an embodiment of an electronic device for realizing all or part of the contents in the one-stop artificial intelligent image processing model construction method, which specifically comprises the following contents:

A processor (processor), a memory (memory), a communication interface (CommunicationsInterface), and a bus; the processor, the memory and the communication interface complete communication with each other through the bus; the communication interface is used for realizing information transmission between the one-stop artificial intelligent image processing model building device and related equipment such as a core service system, a user terminal, a related database and the like; the logic controller may be a desktop computer, a tablet computer, a mobile terminal, etc., and the embodiment is not limited thereto. In this embodiment, the logic controller may refer to an embodiment of the one-stop type artificial intelligence image processing model construction method and an embodiment of the one-stop type artificial intelligence image processing model construction device, and the contents thereof are incorporated herein and are not repeated here.

It is understood that the user terminal may include a smart phone, a tablet electronic device, a network set top box, a portable computer, a desktop computer, a Personal Digital Assistant (PDA), a vehicle-mounted device, a smart wearable device, etc. Wherein, intelligent wearing equipment can include intelligent glasses, intelligent wrist-watch, intelligent bracelet etc..

In practical applications, part of the one-stop artificial intelligence image processing model construction method may be executed on the electronic device side as described above, or all operations may be completed in the client device. Specifically, the selection may be made according to the processing capability of the client device, and restrictions of the use scenario of the user. The application is not limited in this regard. If all operations are performed in the client device, the client device may further include a processor.

The client device may have a communication module (i.e. a communication unit) and may be connected to a remote server in a communication manner, so as to implement data transmission with the server. The server may include a server on the side of the task scheduling center, and in other implementations may include a server of an intermediate platform, such as a server of a third party server platform having a communication link with the task scheduling center server. The server may include a single computer device, a server cluster formed by a plurality of servers, or a server structure of a distributed device.

Fig. 9 is a schematic block diagram of a system configuration of an electronic device 9600 according to an embodiment of the present application. As shown in fig. 9, the electronic device 9600 may include a central processor 9100 and a memory 9140; the memory 9140 is coupled to the central processor 9100. Notably, this fig. 9 is exemplary; other types of structures may also be used in addition to or in place of the structures to implement telecommunications functions or other functions.

In one embodiment, the one-stop artificial intelligence image processing model building method functionality may be integrated into the central processor 9100. The central processor 9100 may be configured to perform the following control:

As can be seen from the above description, in the electronic device provided by the embodiment of the present application, by creating an image labeling task, determining a structure of a corresponding image processing model, and training the image processing model according to image data stored in the standardized label, so as to obtain the image processing model; according to the model definition specification corresponding to the image processing model, a corresponding preprocessing interface, an inference interface and a post-processing interface of the target deployment hardware are determined, image data stored by a standardized label processed by the preprocessing interface is input to the image processing model through the inference interface, model output of the image processing model is processed through the post-processing interface, corresponding model precision is determined according to the model output, and the image processing model is deployed to the target deployment hardware when the model precision meets the preset precision condition, so that the integrated integration of full flows of data labeling, model training, model conversion, model deployment and the like can be realized, development efficiency is improved, development cost is reduced, and high consistency of training precision and deployment precision is ensured.

In another embodiment, the one-stop artificial intelligent image processing model building apparatus may be configured separately from the cpu 9100, for example, the one-stop artificial intelligent image processing model building apparatus may be configured as a chip connected to the cpu 9100, and the one-stop artificial intelligent image processing model building method function is implemented under the control of the cpu.

As shown in fig. 9, the electronic device 9600 may further include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It is noted that the electronic device 9600 need not include all of the components shown in fig. 9; in addition, the electronic device 9600 may further include components not shown in fig. 9, and reference may be made to the related art.

As shown in fig. 9, the central processor 9100, sometimes referred to as a controller or operational control, may include a microprocessor or other processor device and/or logic device, which central processor 9100 receives inputs and controls the operation of the various components of the electronic device 9600.

The memory 9140 may be, for example, one or more of a buffer, a flash memory, a hard drive, a removable media, a volatile memory, a non-volatile memory, or other suitable device. The information about failure may be stored, and a program for executing the information may be stored. And the central processor 9100 can execute the program stored in the memory 9140 to realize information storage or processing, and the like.

The input unit 9120 provides input to the central processor 9100. The input unit 9120 is, for example, a key or a touch input device. The power supply 9170 is used to provide power to the electronic device 9600. The display 9160 is used for displaying display objects such as images and characters. The display may be, for example, but not limited to, an LCD display.

The memory 9140 may be a solid state memory such as Read Only Memory (ROM), random Access Memory (RAM), SIM card, etc. But also a memory which holds information even when powered down, can be selectively erased and provided with further data, an example of which is sometimes referred to as EPROM or the like. The memory 9140 may also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 may include an application/function storage portion 9142, the application/function storage portion 9142 storing application programs and function programs or a flow for executing operations of the electronic device 9600 by the central processor 9100.

The memory 9140 may also include a data store 9143, the data store 9143 for storing data, such as contacts, digital data, pictures, sounds, and/or any other data used by an electronic device. The driver storage portion 9144 of the memory 9140 may include various drivers of the electronic device for communication functions and/or for performing other functions of the electronic device (e.g., messaging applications, address book applications, etc.).

The communication module 9110 is a transmitter/receiver that transmits and receives signals via the antenna 9111. The communication module 9110 (transmitter/receiver) is coupled to the central processor 9100 to provide input signals and receive output signals, as in the case of conventional mobile communication terminals.

Based on different communication technologies, a plurality of communication modules 9110, such as a cellular network module, a bluetooth module, and/or a wireless local area network module, etc., may be provided in the same electronic device. The communication module 9110 (transmitter/receiver) is also coupled to a speaker 9131 and a microphone 9132 via an audio processor 9130 to provide audio output via the speaker 9131 and to receive audio input from the microphone 9132 to implement usual telecommunications functions. The audio processor 9130 can include any suitable buffers, decoders, amplifiers and so forth. In addition, the audio processor 9130 is also coupled to the central processor 9100 so that sound can be recorded locally through the microphone 9132 and sound stored locally can be played through the speaker 9131.

An embodiment of the present application further provides a computer-readable storage medium capable of implementing all steps in the one-stop type artificial intelligence image processing model construction method in which the execution subject is a server or a client, the computer-readable storage medium storing thereon a computer program which, when executed by a processor, implements all steps in the one-stop type artificial intelligence image processing model construction method in which the execution subject is a server or a client, for example, the processor implements the following steps when executing the computer program:

As can be seen from the above description, the computer readable storage medium provided by the embodiments of the present application determines the structure of a corresponding image processing model by creating an image labeling task, and performs training of the image processing model according to the image data stored in the standardized label, so as to obtain the image processing model; according to the model definition specification corresponding to the image processing model, a corresponding preprocessing interface, an inference interface and a post-processing interface of the target deployment hardware are determined, image data stored by a standardized label processed by the preprocessing interface is input to the image processing model through the inference interface, model output of the image processing model is processed through the post-processing interface, corresponding model precision is determined according to the model output, and the image processing model is deployed to the target deployment hardware when the model precision meets the preset precision condition, so that the integrated integration of full flows of data labeling, model training, model conversion, model deployment and the like can be realized, development efficiency is improved, development cost is reduced, and high consistency of training precision and deployment precision is ensured.

Embodiments of the present application also provide a computer program product capable of implementing all the steps in the one-stop artificial intelligence image processing model construction method in which the execution subject is a server or a client, and the computer program/instructions implement the steps of the one-stop artificial intelligence image processing model construction method when executed by a processor, for example, the computer program/instructions implement the steps of:

As can be seen from the above description, the computer program product provided by the embodiments of the present application determines the structure of the corresponding image processing model by creating an image labeling task, and performs training of the image processing model according to the image data stored in the standardized label, so as to obtain the image processing model; according to the model definition specification corresponding to the image processing model, a corresponding preprocessing interface, an inference interface and a post-processing interface of the target deployment hardware are determined, image data stored by a standardized label processed by the preprocessing interface is input to the image processing model through the inference interface, model output of the image processing model is processed through the post-processing interface, corresponding model precision is determined according to the model output, and the image processing model is deployed to the target deployment hardware when the model precision meets the preset precision condition, so that the integrated integration of full flows of data labeling, model training, model conversion, model deployment and the like can be realized, development efficiency is improved, development cost is reduced, and high consistency of training precision and deployment precision is ensured.

It will be apparent to those skilled in the art that embodiments of the present invention may be provided as a method, apparatus, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

The principles and embodiments of the present invention have been described in detail with reference to specific examples, which are provided to facilitate understanding of the method and core ideas of the present invention; meanwhile, as those skilled in the art will have variations in the specific embodiments and application scope in accordance with the ideas of the present invention, the present description should not be construed as limiting the present invention in view of the above.

Claims

1. A method for constructing a one-stop artificial intelligence image processing model, the method comprising:

2. The method for constructing a one-stop artificial intelligence image processing model according to claim 1, wherein the determining a corresponding pre-labeling model according to the image labeling task and performing pre-labeling processing on preset image data according to the pre-labeling model comprises:

3. The method for constructing a one-stop artificial intelligence image processing model according to claim 2, wherein the determining a corresponding pre-labeling model according to the task type of the image labeling task and the data characteristics of the preset image data comprises:

4. A method of constructing a one-stop artificial intelligence image processing model according to claim 3, wherein the matching the corresponding pre-labeling model from a pre-set model library according to the image feature and the task type comprises:

5. The method for constructing a one-stop artificial intelligence image processing model according to claim 1, wherein the storing the image data subjected to the pre-labeling processing according to the data type by a corresponding standardized label comprises:

6. The method for constructing a one-stop artificial intelligence image processing model according to claim 1, wherein the converting the image processing model into a corresponding executable format and quantizing floating point parameters of the image processing model into integer numbers or fixed point numbers according to the specific type of the target deployment hardware to obtain the image processing model after format conversion and parameter quantization comprises:

7. The one-stop artificial intelligence image processing model construction method according to claim 1, wherein the inputting the image data stored in the standardized tag processed by the preprocessing interface to the image processing model through the reasoning interface, and processing the model output of the image processing model through the post-processing interface, determining the corresponding model accuracy according to the model output, comprises:

8. A one-stop artificial intelligence image processing model construction apparatus, the apparatus comprising:

9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor performs the steps of the one-stop artificial intelligence image processing model building method of any one of claims 1 to 7 when the program is executed.

10. A computer readable storage medium having stored thereon a computer program, which when executed by a processor, implements the steps of the one-stop artificial intelligence image processing model construction method of any of claims 1 to 7.