AI algorithm service platform construction method based on deep learning network

Abstract

The invention relates to the technical field of computer artificial intelligence, and discloses an AI algorithm service platform construction method based on a deep learning network, and the method comprises the steps: constructing an initial service platform; constructing a target statistical algorithm, and processing image input data by adopting a Yolox network model of an improved prediction end; constructing a target segmentation algorithm, and performing target detection according to the target number of the image input data; constructing a target recognition algorithm, and recognizing the target image region based on a preset convolutional neural network; and adding the target statistical algorithm, the target segmentation algorithm and the target recognition algorithm into the initial service platform, and determining a virtualized resource pool and an application layer to complete construction of the AI algorithm service platform based on the deep learning network. The AI algorithm service platform comprising the target statistical algorithm, the target segmentation algorithm and the target recognition algorithm is constructed, and the method has the advantages that the algorithm convergence speed is high, the parameter and floating point operand calculation amount is reduced, and the transmission speed and performance are better.

Description

Technical Field

[0001] The present invention relates to the field of computer artificial intelligence technology, and in particular to a method for constructing an AI algorithm service platform based on a deep learning network. Background Art

[0002] With the rapid development of computer artificial intelligence technology, computer vision technology has gradually penetrated into all walks of life with its great convenience. However, due to geographical and production restrictions, specific fields still face the problem of difficulty in applying and promoting high-tech, resulting in low production efficiency and backward technology. Similarly, deep learning is a new research direction in the field of machine learning and has been widely used in the industrial field. Therefore, how to apply the relevant knowledge and technology of deep learning in specific fields to achieve changes and improvements in production efficiency.

[0003] With the development of sensor technology and image processing technology, a large number of regions have installed monitoring equipment for real-time and effective dynamic monitoring. The effective use of these images can achieve the purpose of refined management to a certain extent. As an emerging machine learning method, deep learning has shown good performance in image processing. The application of this method is bound to improve management efficiency.

[0004] Taking the advanced single-stage target detection framework Yolo as an example, its detection head is always coupled, which greatly affects the convergence performance of the detection algorithm. The original feature pyramid network FPN used by the convolutional neural network CNN uses a top-down approach to transmit high-level features. This processing method does not take into account the enhancement of the underlying edges, shapes and other features. In reality, there are usually poor lighting conditions and the monitoring targets squeeze and occlude each other, so the edges of the target contours in the collected images are not clear enough. Summary of the invention

[0005] The present invention provides a method for constructing an AI algorithm service platform based on a deep learning network, and constructs an AI algorithm service platform including a target statistical algorithm, a target segmentation algorithm and a target recognition algorithm. The method has the advantages of fast algorithm convergence speed, reduced parameter and floating-point operation calculation amount, better transmission speed and performance, and improves monitoring effect and management level.

[0006] The present invention provides a method for constructing an AI algorithm service platform based on a deep learning network, comprising:

[0007] Build the initial service platform;

[0008] Constructing a target statistical algorithm for processing image input data using a Yolox network model of an improved prediction end to perform target statistics on the image input data to obtain a target number of the image input data;

[0009] Constructing a target segmentation algorithm for performing target detection according to the number of targets in the image input data, so as to complete the segmentation and contour extraction of the target and obtain the target image area;

[0010] Constructing a target recognition algorithm for identifying the target image area based on a preset convolutional neural network to obtain a detection result of the original image data;

[0011] The target statistics algorithm, target segmentation algorithm and target recognition algorithm are added to the initial service platform, and the virtualized resource pool and application layer of the AI ​​algorithm service platform are determined to complete the construction of the AI ​​algorithm service platform based on the deep learning network.

[0012] Furthermore, in the step of constructing a target statistical algorithm for processing image input data using the Yolox network model of the improved prediction end to perform target statistics on the image input data and obtain the number of targets of the image input data, target detection specifically includes:

[0013] Using the Anchor-free method, under the premise of pre-setting the standard point range, the center of the detection target is set as the positive sample, the height and width of the prediction box are set, and the FPN level of the feature pyramid structure of each object is determined; the Multi-positives strategy is adopted, and when selecting the candidate box of the target box, the number of selections is greater than 1 to optimize the prediction and bring beneficial gradients. Finally, the SimOTA algorithm is used to obtain three prediction results for each feature layer:

[0014] Realize the prediction of the target box position: use the Reg_output(h,w,4) function to calculate the center point position of the prediction box, the height and width of the prediction box;

[0015] Complete the judgment of the foreground and background of the target box: use the Sigmoid function to process Obj_output(h,w,1) and calculate the confidence of the foreground of the target box;

[0016] Determine the target type: After processing Cls_output(h,w,num_classes) with the Sigmoid function, set a score for each type of target and get the confidence of each type of target.

[0017] Furthermore, the step of constructing a target statistical algorithm for processing image input data using the Yolox network model of the improved prediction end to perform target statistics on the image input data and obtain the target number of the image input data, before target detection, further includes:

[0018] Collect or receive original image data, and perform data preprocessing on the original image data to obtain the image input data; wherein the data preprocessing includes geometric correction, image fusion, image mosaicking, and image cropping.

[0019] Furthermore, the step of constructing a target segmentation algorithm for performing target detection according to the number of targets in the image input data to complete target segmentation and contour extraction and obtain a target image area includes:

[0020] The MaskR-CNN model is used to preprocess and extract features of the original image collected by the image acquisition device to obtain features of the image input data and obtain corresponding feature maps;

[0021] Based on the feature maps obtained in the above steps, the RoI area of ​​each image is extracted using the region proposal network RP, the foreground and background are binary classified, and the bounding box regression and non-maximum suppression methods are used for calculation and processing;

[0022] In the RoIAlign layer, the results calculated in the previous step are unified in size;

[0023] The calculation results of the previous step are passed through branches including fully connected layers and fully convolutional neural networks to perform target classification, bounding box regression operations and pixel segmentation.

[0024] Furthermore, in the step of constructing a target recognition algorithm for identifying the target image area based on a preset convolutional neural network to obtain the detection result of the original image data, the preset convolutional neural network includes a feature extraction network and a classification and positioning sub-network;

[0025] The feature extraction network extracts representative features from the target image area through stacked convolution layers, pooling layers and activation functions, and the features can capture key information in the data; the classification and positioning subnetwork maps the extracted features to specific categories or positions through fully connected layers and softmax functions.

[0026] Furthermore, the feature extraction network uses a deep residual network as a basic structure, and combines a feature pyramid network to extract multi-scale image features from the target image region, and image features of different scales are used to detect objects of interest of different scales;

[0027] The deep residual network adopts the ResNet-50 structure to extract mid- and high-level semantic features and introduces residual learning to prevent network overfitting. The feature pyramid network constructs a 5-layer feature pyramid based on the output of the 3rd to 5th residual modules of the deep residual network. These feature pyramids represent downsampled feature maps with step sizes of {8, 16, 32, 64, 128} respectively.

[0028] Furthermore, the classification and positioning subnetwork takes the feature map output by the feature extraction network as input, and predicts the category probability for each position in the feature map through a convolutional layer;

[0029] Each prediction is a rectangular box of different scales and different aspect ratios centered at a certain position in the feature map. The rectangular box is called an anchor point. For each anchor point, the classification subnetwork outputs a classification result related to the category. If the number of target recognition categories is K, its output is a vector of length K, in which each element represents the probability that the image area where the anchor point is located belongs to a specific category.

[0030] The present invention also provides an AI algorithm service platform construction device based on a deep learning network, comprising:

[0031] The first building module is used to build an initial service platform;

[0032] The second construction module is used to construct a target statistical algorithm, which is used to process the image input data using the Yolox network model of the improved prediction end to perform target statistics on the image input data to obtain the target number of the image input data;

[0033] The third construction module is used to construct a target segmentation algorithm, which is used to perform target detection according to the number of targets in the image input data, so as to complete the segmentation and contour extraction of the target and obtain the target image area;

[0034] A fourth construction module is used to construct a target recognition algorithm, which is used to recognize the target image area based on a preset convolutional neural network to obtain a detection result of the original image data;

[0035] A determination module is used to add the target statistical algorithm, target segmentation algorithm and target recognition algorithm to the initial service platform, and determine the virtualized resource pool and application layer of the AI ​​algorithm service platform to complete the construction of the AI ​​algorithm service platform based on the deep learning network.

[0036] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0037] The present invention also provides a computer-readable storage medium on which a computer program is stored, and the computer program implements the steps of the above method when executed by a processor.

[0038] The beneficial effects of the present invention are:

[0039] The present invention constructs an AI algorithm service platform, including the construction of a target statistical algorithm, a target segmentation algorithm, a target recognition algorithm and a virtualized resource pool. The target detection algorithm uses an improved Yolox network model of the prediction end to process image input data; the target segmentation algorithm performs target detection according to the number of targets in the image input data to complete target segmentation and contour extraction; the target recognition algorithm recognizes the target image area based on a preset convolutional neural network; finally, the virtualized resource pool and application layer of the AI ​​algorithm service platform are determined to provide a hardware foundation for the construction and training of the algorithm. The algorithm platform has the advantages of fast algorithm convergence speed, reduced calculation amount of parameters and floating-point operations, better transmission speed and performance, and improves the monitoring effect and management level. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 The figure is a schematic diagram of a method flow according to an embodiment of the present invention.

[0041] Figure 2 FIG. 1 is a schematic diagram of a device structure according to an embodiment of the present invention.

[0042] Figure 3 The figure is a schematic diagram of the internal structure of a computer device according to an embodiment of the present invention.

[0043] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0044] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0045] like Figure 1 As shown, the present invention provides a method for constructing an AI algorithm service platform based on a deep learning network, comprising:

[0046] S1. Build the initial service platform;

[0047] S2. constructing a target statistical algorithm for processing the image input data by using the Yolox network model of the improved prediction end to perform target statistics on the image input data to obtain a target number of the image input data;

[0048] S21, collecting or receiving original image data, and performing data preprocessing on the original image data to obtain the image input data; wherein the data preprocessing includes geometric correction, image fusion, image mosaicking, and image cropping.

[0049] Among them, geometric correction includes geometric precision correction, image registration, orthorectification, etc. Image fusion: merge multiple images into a single image to enhance the expression of information. Image mosaicking and cropping: crop the image as needed to remove unnecessary parts.

[0050] The specific methods of image preprocessing include:

[0051] Denoising: Remove noise from the image to improve image quality. Enhancement: Enhance the contrast and clarity of the image through grayscale transformation, histogram correction and other methods. Smoothing: Smooth the image to reduce the interference of noise and details. Transformation: Change the size and angle of the image to meet the needs of subsequent processing. Grayscale and binarization: Convert color images to grayscale images or binary images to simplify data processing.

[0052] The purpose of image preprocessing is to eliminate irrelevant information in the image, restore useful real information, enhance the detectability of relevant information, and simplify the data to the maximum extent, thereby improving the reliability of feature extraction, image segmentation, matching and recognition.

[0053] S22. Improve and enhance the performance of Yolox using Anchor-free (no anchor node required), Multi-positives (multiple positive samples) and SimOTA (matching strategy) methods. The specific operations are as follows:

[0054] In order to improve the detection speed, the Anchor-free method is used to reduce the number of predictions in the collected image. Under the premise of pre-setting the standard point range, the center of the detection target is set as the positive sample, the height and width of the prediction box are set, and finally the FPN level of the feature pyramid structure of each object is determined. However, Anchor-free also has a disadvantage when processing images, that is, when it sets the positive sample in the image, it only selects one positive sample. This processing method will cause the number of positive samples to be unable to match the number of negative samples. In order to solve this problem, the Multi-positives strategy is adopted. When selecting the candidate box of the target box, the number of selections is greater than 1, and those high-quality predictions are optimized to bring beneficial gradients. This processing method can alleviate the extreme imbalance of positive and negative sampling during training. Finally, considering the real-time problem of the entire platform construction, SimOTA is an algorithm that removes the process of solving the optimal solution of Sinkhorn-Knopp iteration in OTA. This processing method can save training time and improve the efficiency of target detection.

[0055] Completing the above operations can reduce the number of channels of each feature map, improve the efficiency of classification and regression, and obtain three prediction results for each feature layer, that is, the detection of the target specifically includes:

[0056] Realize the prediction of the target box position: use the Reg_output(h,w,4) function to calculate the center point position of the prediction box, the height and width of the prediction box;

[0057] Complete the judgment of the foreground and background of the target box: use the Sigmoid function to process Obj_output(h,w,1) and calculate the confidence of the foreground of the target box;

[0058] Determine the target type: After processing Cls_output(h,w,num_classes) with the Sigmoid function, set a score for each type of target and get the confidence of each type of target.

[0059] The above improved Yolox model improves the accuracy and real-time performance of target detection.

[0060] S3, constructing a target segmentation algorithm for performing target detection according to the number of targets in the image input data, so as to complete the segmentation and contour extraction of the target and obtain the target image area;

[0061] Image segmentation is used to separate the foreground and background of an object, and Mask R-CNN is used as the benchmark algorithm model. The Mask R-CNN network model realizes the combination of target detection and image segmentation. In the actual operation process, the image input data is first input into a pre-trained neural network to obtain the corresponding feature map; then a predetermined number of ROIs are set for each point in the feature map to obtain multiple candidate ROIs; then these candidate ROIs are sent to the RPN network for binary classification (foreground or background) and BB regression, and some candidate ROIs are filtered out; then the remaining ROIs are ROIAligned; finally, these ROIs are classified, BB regression and MASK generation are performed.

[0062] The MaskR-CNN model is improved by applying Swin-Transform to the model for feature extraction. At the same time, a bottom-up enhancement path is added after the original feature pyramid network FPN. This improvement can improve the positioning accuracy of the bounding box and solve the problems of high density and severe occlusion. The specific steps are:

[0063] S31, using the MaskR-CNN model to preprocess and extract features of the original image collected by the image acquisition device to obtain features of the image input data and obtain a corresponding feature map;

[0064] S32, based on the feature map obtained in the above steps, the RoI region of each image is extracted using the region proposal network RP, the foreground and background are binary classified, and the bounding box regression and non-maximum suppression methods are used for calculation and processing;

[0065] S33, unifying the size of the result calculated in the previous step in the RoIAlign layer;

[0066] S34, the calculation result of the previous step is used for target classification, bounding box regression operation and pixel segmentation through branches including a fully connected layer and a fully convolutional neural network.

[0067] S4, constructing a target recognition algorithm for identifying the target image area based on a preset convolutional neural network to obtain a detection result of the original image data;

[0068] The feature extraction network is the first step of the network structure. Through stacked convolutional layers, pooling layers, activation functions and other components, representative features are extracted from the original data. These features can capture the key information in the data, allowing the model to better understand and express the input data. The design and optimization of the feature extraction network directly affects the performance and generalization ability of the network, and plays a vital role in tasks such as image recognition and target detection. At the same time, the classification and positioning subnetwork constitutes the subsequent part of the network structure. Through a series of fully connected layers and softmax functions and other components, the extracted features are mapped to specific categories or locations. The classification and positioning subnetwork can accurately classify and locate the input data by learning the association rules in the data and the relationship between the features, and realize accurate recognition of different categories and locations.

[0069] Feature extraction network:

[0070] The feature extraction network uses a deep residual network as a basic structure, and combines a feature pyramid network to extract multi-scale image features from the target image area, and image features of different scales are used to detect objects of interest of different scales;

[0071] The deep residual network uses the ResNet-50 structure to extract mid- and high-level semantic features, and introduces residual learning to prevent network overfitting. The number of layers of the deep residual network can be adjusted according to the actual task and the scale of training data, so as to strike a balance between computational efficiency and detection accuracy. The feature pyramid network builds a 5-layer feature pyramid based on the output of the 3rd to 5th residual modules of the deep residual network. These feature pyramids represent downsampled feature maps with step sizes of {8, 16, 32, 64, 128}, respectively. The feature pyramid enables the system to effectively handle the challenge of large scale changes of targets on the battlefield as the detection distance increases. At the same time, it also integrates the problem of multi-scale detection into the learning process of the entire network, which has greater flexibility.

[0072] Classification and positioning sub-network:

[0073] The classification and localization subnetwork adopts the idea based on anchor points. It takes the feature map output by the feature extraction network as input and predicts the category probability for each position in the feature map through several convolutional layers. Each prediction is a rectangular box with different scales and different aspect ratios centered at a certain position in the feature map. Since the relative offset of each predicted rectangular box to its center position is fixed, this predicted rectangular box is called an anchor point. For each anchor point, the classification subnetwork outputs the classification result related to the category. If the number of target recognition categories is K, its output is a vector of length K, in which each element represents the probability that the image area where the anchor point is located belongs to a specific category. The localization subnetwork outputs a vector of length 4, which represents the offset of the anchor point relative to the true target rectangular box in the horizontal and vertical coordinates and the four dimensions of length and width. If the number of anchor points is A, for a feature map of size W×H, the output of the classification subnetwork is a tensor of W×H×KA, and the output of the localization subnetwork is a tensor of W×H×4A. The final detection result can be obtained after corresponding post-processing.

[0074] S5. Add the target statistical algorithm, target segmentation algorithm and target recognition algorithm to the initial service platform, and determine the virtualized resource pool and application layer of the AI ​​algorithm service platform to complete the construction of the AI ​​algorithm service platform based on the deep learning network.

[0075] As a key element in realizing the converged infrastructure structure, the virtual resource pool is a collection of shared servers, storage and networks, which can be reconfigured faster according to the requirements of the application, so that administrators can support changes in business needs more easily and quickly than before. In a cloud computing environment, resources are no longer scattered hardware, but physical servers are integrated to form one or more logical virtual resource pools, sharing computing, storage and network resources. The resource pool can delegate control of host (or cluster) resources, and its advantages are very obvious when using the resource pool to divide all resources within the cluster. Multiple resource pools can be created as direct children of the host or cluster and configured. When building a virtualized resource pool for the algorithm service platform of the present invention, the deployment of the framework is achieved by adopting the Kubernetes container orchestration engine and the Torchserve deep learning model. Based on this set of hardware resources, target detection and segmentation algorithms can be completed and improved therein. The application layer mainly implements a series of tasks that meet actual needs, such as status monitoring and data output.

[0076] Finally, in order to achieve the scalability of the designed platform, the design of the interface was considered, including two interfaces. One is the RESTFUL interface for external communication, which is mainly used for image transmission and storage services and to implement the calling function of the algorithm model; the other interface is mainly used for model calling, which will mainly deploy the service calling cluster through the Kubernetes container orchestration engine, and then use the Torchserve deep learning model deployment framework to deploy the called interface of the corresponding algorithm model.

[0077] The present invention constructs an AI algorithm service platform, including the construction of a target statistical algorithm, a target segmentation algorithm, a target recognition algorithm and a virtualized resource pool. The target detection algorithm uses an improved Yolox network model of the prediction end to process image input data; the target segmentation algorithm performs target detection according to the number of targets in the image input data to complete target segmentation and contour extraction; the target recognition algorithm recognizes the target image area based on a preset convolutional neural network; finally, the virtualized resource pool and application layer of the AI ​​algorithm service platform are determined to provide a hardware foundation for the construction and training of the algorithm. The algorithm platform has the advantages of fast algorithm convergence speed, reduced calculation amount of parameters and floating-point operations, better transmission speed and performance, and improves the monitoring effect and management level.

[0078] like Figure 2 As shown, the present invention also provides an AI algorithm service platform construction device based on a deep learning network, comprising:

[0079] The first building block 1 is used to build an initial service platform;

[0080] The second construction module 2 is used to construct a target statistical algorithm, which is used to process the image input data using the Yolox network model of the improved prediction end to perform target statistics on the image input data to obtain the target number of the image input data;

[0081] The third construction module 3 is used to construct a target segmentation algorithm, which is used to perform target detection according to the number of targets in the image input data, so as to complete the segmentation and contour extraction of the target and obtain the target image area;

[0082] The fourth construction module 4 is used to construct a target recognition algorithm, which is used to recognize the target image area based on a preset convolutional neural network to obtain a detection result of the original image data;

[0083] Determine module 5, which is used to add the target statistical algorithm, target segmentation algorithm and target recognition algorithm to the initial service platform, and determine the virtualized resource pool and application layer of the AI ​​algorithm service platform to complete the construction of the AI ​​algorithm service platform based on the deep learning network.

[0084] In one embodiment, in the second building module 2, target detection specifically includes:

[0085] Using the Anchor-free method, under the premise of pre-setting the standard point range, the center of the detection target is set as the positive sample, the height and width of the prediction box are set, and the FPN level of the feature pyramid structure of each object is determined; the Multi-positives strategy is adopted, and when selecting the candidate box of the target box, the number of selections is greater than 1 to optimize the prediction and bring beneficial gradients. Finally, the SimOTA algorithm is used to obtain three prediction results for each feature layer:

[0086] Realize the prediction of the target box position: use the Reg_output(h,w,4) function to calculate the center point position of the prediction box, the height and width of the prediction box;

[0087] Complete the judgment of the foreground and background of the target box: use the Sigmoid function to process Obj_output(h,w,1) and calculate the confidence of the foreground of the target box;

[0088] Determine the target type: After processing Cls_output(h,w,num_classes) with the Sigmoid function, set a score for each type of target and get the confidence of each type of target.

[0089] In one embodiment, the second building module 2, before the target detection, further includes:

[0090] Collect or receive original image data, and perform data preprocessing on the original image data to obtain the image input data; wherein the data preprocessing includes geometric correction, image fusion, image mosaicking, and image cropping.

[0091] In one embodiment, the third building block 3 includes:

[0092] An extraction unit, used to use a MaskR-CNN model to preprocess and extract features from an original image acquired by an image acquisition device, so as to obtain features of the image input data and acquire a corresponding feature map;

[0093] A processing unit is used to extract the RoI region of each image based on the feature map obtained in the above steps using the region proposal network RP, perform binary classification on the foreground and background, and perform calculation processing using bounding box regression and non-maximum suppression methods;

[0094] The unification unit is used to unify the size of the result calculated in the previous step in the RoIAlign layer;

[0095] The classification unit is used to perform target classification, bounding box regression and pixel segmentation on the calculation results of the previous step through branches including fully connected layers and fully convolutional neural networks.

[0096] In one embodiment, in the third building block 3, the preset convolutional neural network includes a feature extraction network and a classification and positioning sub-network;

[0097] The feature extraction network extracts representative features from the target image area through stacked convolution layers, pooling layers and activation functions, and the features can capture key information in the data; the classification and positioning subnetwork maps the extracted features to specific categories or positions through fully connected layers and softmax functions.

[0098] In one embodiment, in the third building module 3, the feature extraction network uses a deep residual network as a basic structure, and combines a feature pyramid network to extract multi-scale image features from the target image area, and image features of different scales are used to detect objects of interest of different scales;

[0099] The deep residual network adopts the ResNet-50 structure to extract mid- and high-level semantic features and introduces residual learning to prevent network overfitting. The feature pyramid network constructs a 5-layer feature pyramid based on the output of the 3rd to 5th residual modules of the deep residual network. These feature pyramids represent downsampled feature maps with step sizes of {8, 16, 32, 64, 128} respectively.

[0100] In one embodiment, in the third building module 3, the classification and positioning subnetwork takes the feature map output by the feature extraction network as input, and predicts the category probability for each position in the feature map through a convolutional layer;

[0101] Each prediction is a rectangular box of different scales and different aspect ratios centered at a certain position in the feature map. The rectangular box is called an anchor point. For each anchor point, the classification subnetwork outputs a classification result related to the category. If the number of target recognition categories is K, its output is a vector of length K, in which each element represents the probability that the image area where the anchor point is located belongs to a specific category.

[0102] The above modules and units are used to execute the corresponding steps in the above-mentioned method for constructing an AI algorithm service platform based on a deep learning network. The specific implementation method is described in the above-mentioned method embodiment and will not be repeated here.

[0103] like Figure 3 As shown, the present invention also provides a computer device, which can be a server, and its internal structure can be as shown in Figure 3 As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store all data required for the process of the method for constructing an AI algorithm service platform based on a deep learning network. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the method for constructing an AI algorithm service platform based on a deep learning network is implemented.

[0104] Those skilled in the art will understand that Figure 3 The structure shown in is merely a block diagram of a portion of the structure related to the present application solution and does not constitute a limitation on the computer device to which the present application solution is applied.

[0105] An embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, any one of the above-mentioned methods for constructing an AI algorithm service platform based on a deep learning network is implemented.

[0106] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing related hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0107] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0108] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for constructing an AI algorithm service platform based on a deep learning network, characterized in that: include: Build the initial service platform; Constructing a target statistical algorithm for processing image input data using a Yolox network model of an improved prediction end to perform target statistics on the image input data to obtain a target number of the image input data; Constructing a target segmentation algorithm for performing target detection according to the number of targets in the image input data, so as to complete the segmentation and contour extraction of the target and obtain the target image area; Constructing a target recognition algorithm for identifying the target image area based on a preset convolutional neural network to obtain a detection result of the original image data; The target statistics algorithm, target segmentation algorithm and target recognition algorithm are added to the initial service platform, and the virtualized resource pool and application layer of the AI ​​algorithm service platform are determined to complete the construction of the AI ​​algorithm service platform based on the deep learning network.

2. The method for constructing an AI algorithm service platform based on a deep learning network according to claim 1, characterized in that: The target statistical algorithm is constructed to process the image input data using the Yolox network model of the improved prediction end to perform target statistics on the image input data to obtain the target number of the image input data. In the step, target detection specifically includes: Using the Anchor-free method, under the premise of pre-setting the standard point range, the center of the detection target is set as the positive sample, the height and width of the prediction box are set, and the FPN level of the feature pyramid structure of each object is determined; the Multi-positives strategy is adopted, and when selecting the candidate box of the target box, the number of selections is greater than 1 to optimize the prediction and bring beneficial gradients. Finally, the SimOTA algorithm is used to obtain three prediction results for each feature layer: Realize the prediction of the target box position: use the Reg_output(h,w,4) function to calculate the center point position of the prediction box, the height and width of the prediction box; Complete the judgment of the foreground and background of the target box: use the Sigmoid function to process Obj_output(h,w,1) and calculate the confidence of the foreground of the target box; Determine the target type: After processing Cls_output(h,w,num_classes) with the Sigmoid function, set a score for each type of target and get the confidence of each type of target.

3. The method for constructing an AI algorithm service platform based on a deep learning network according to claim 2, characterized in that: The step of constructing a target statistics algorithm for processing image input data using the Yolox network model of the improved prediction end to perform target statistics on the image input data and obtain the target number of the image input data, before target detection, further includes: Collect or receive original image data, and perform data preprocessing on the original image data to obtain the image input data; wherein the data preprocessing includes geometric correction, image fusion, image mosaicking, and image cropping.

4. The method for constructing an AI algorithm service platform based on a deep learning network according to claim 1, characterized in that: The step of constructing a target segmentation algorithm for performing target detection according to the number of targets in the image input data to complete target segmentation and contour extraction and obtain a target image area includes: The MaskR-CNN model is used to preprocess and extract features of the original image collected by the image acquisition device to obtain features of the image input data and obtain corresponding feature maps; Based on the feature maps obtained in the above steps, the RoI area of ​​each image is extracted using the region proposal network RP, the foreground and background are binary classified, and the bounding box regression and non-maximum suppression methods are used for calculation and processing; In the RoIAlign layer, the results calculated in the previous step are unified in size; The calculation results of the previous step are passed through branches including fully connected layers and fully convolutional neural networks to perform target classification, bounding box regression operations and pixel segmentation.

5. The method for constructing an AI algorithm service platform based on a deep learning network according to claim 1, characterized in that: The target recognition algorithm is constructed to identify the target image area based on a preset convolutional neural network to obtain the detection result of the original image data, wherein the preset convolutional neural network includes a feature extraction network and a classification and positioning sub-network; The feature extraction network extracts representative features from the target image area through stacked convolution layers, pooling layers and activation functions, and the features can capture key information in the data; the classification and positioning subnetwork maps the extracted features to specific categories or positions through fully connected layers and softmax functions.

6. The method for constructing an AI algorithm service platform based on a deep learning network according to claim 5, characterized in that: The feature extraction network uses a deep residual network as a basic structure, and combines a feature pyramid network to extract multi-scale image features from the target image area, and image features of different scales are used to detect objects of interest of different scales; The deep residual network adopts the ResNet-50 structure to extract mid- and high-level semantic features and introduces residual learning to prevent network overfitting. The feature pyramid network constructs a 5-layer feature pyramid based on the output of the 3rd to 5th residual modules of the deep residual network. These feature pyramids represent downsampled feature maps with step sizes of {8, 16, 32, 64, 128} respectively.

7. The method for constructing an AI algorithm service platform based on a deep learning network according to claim 6, characterized in that: The classification and positioning subnetwork takes the feature map output by the feature extraction network as input, and predicts the category probability for each position in the feature map through the convolution layer; Each prediction is a rectangular box of different scales and different aspect ratios centered at a certain position in the feature map. The rectangular box is called an anchor point. For each anchor point, the classification subnetwork outputs a classification result related to the category. If the number of target recognition categories is K, its output is a vector of length K, in which each element represents the probability that the image area where the anchor point is located belongs to a specific category.

8. A device for constructing an AI algorithm service platform based on a deep learning network, characterized in that: include: The first building module is used to build an initial service platform; The second construction module is used to construct a target statistical algorithm, which is used to process the image input data using the Yolox network model of the improved prediction end to perform target statistics on the image input data to obtain the target number of the image input data; The third construction module is used to construct a target segmentation algorithm, which is used to perform target detection according to the number of targets in the image input data, so as to complete the segmentation and contour extraction of the target and obtain the target image area; A fourth construction module is used to construct a target recognition algorithm, which is used to recognize the target image area based on a preset convolutional neural network to obtain a detection result of the original image data; A determination module is used to add the target statistical algorithm, target segmentation algorithm and target recognition algorithm to the initial service platform, and determine the virtualized resource pool and application layer of the AI ​​algorithm service platform to complete the construction of the AI ​​algorithm service platform based on the deep learning network.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images