Image data screening method based on multiple artificial intelligence classification algorithms

Abstract

The present invention relates to the technical field of image data screening, and discloses an image data screening method based on a plurality of artificial intelligence classification algorithms. Classification models are obtained by respectively training a plurality of classification algorithms, and feature analysis is performed on image data to be classified. Suitable classification models are matched based on the characteristics of different classification models, thereby realizing efficient and automatic screening of SMT images and improving the accuracy of screening results.

Description

Technical Field

[0001] The present invention relates to the technical field of image data screening, and in particular to an image data screening method based on multiple artificial intelligence classification algorithms. Background Art

[0002] In the field of electronic manufacturing, surface mount technology (SMT) has become a mainstream process and is widely used in the manufacturing process of various electronic products. With the continuous development and improvement of SMT technology, the requirements for production efficiency and product quality are getting higher and higher. With the continuous expansion and complexity of SMT production lines, more and more image data are generated, and valid data and invalid data are mixed, but not all image data are equally important for production quality control.

[0003] Therefore, it is necessary to screen out valuable image data from massive image data, while the existing SMT image data screening methods are mostly manual screening methods, such as visual inspection, manual classification, etc., which have limitations such as time-consuming, error-prone, and not suitable for large-scale data processing. Currently, there is a lack of efficient, accurate, and automated methods for SMT image data screening on the market. Summary of the invention

[0004] In view of this, the object of the present invention is to provide an image data screening method based on multiple artificial intelligence classification algorithms to achieve efficient and automated screening of SMT images and improve the accuracy of screening results.

[0005] The present invention discloses an image data screening method based on multiple artificial intelligence classification algorithms, the method comprising the following steps:

[0006] Collecting SMT image data, and storing the collected SMT image data in a database as first storage data;

[0007] The first storage data is saved in a structure where the first-level folder is the component type, the second-level folder is the component subdivision model, and the third-level folder is the defect type; wherein the second-level folder is a subfolder of the first-level folder, and the third-level folder is a subfolder of the second-level folder;

[0008] Performing data processing operations on the first stored data, and storing the processed data in a structure where the first-level folder is the component type, the second-level folder is the defect type, and the third-level folder is the component subdivision model as the second stored data;

[0009] Performing a primary screening operation on the second stored data, the primary screening operation comprising screening the second stored data according to useful data and useless data, and saving the data after the primary screening separately in a useful data folder and a useless data folder as the third stored data;

[0010] Constructing initial classification models respectively through different classification algorithms, and training the initial classification models based on the third stored data to obtain classification models corresponding to different classification algorithms; the classification algorithms include AlexNet algorithm, VGGNet-16 algorithm, GoogleNet algorithm and ResNet-18 algorithm;

[0011] A target classification model is selected from different classification models according to the SMT image data to be classified, and the screening result is output through the target classification model.

[0012] Furthermore, constructing an initial classification model based on the AlexNet algorithm includes designing an AlexNet model including local perception field enhancement, introducing multi-scale convolution kernels in some convolution layers, and increasing the size of the convolution kernels.

[0013] Furthermore, constructing an initial classification model based on the VGGNet-16 algorithm includes designing a VGGNet-16 model including deep feature fusion, introducing skip connections in the middle layer of the network, and fusing shallow features with deep features.

[0014] Furthermore, building an initial classification model based on the Google Net algorithm includes designing a Google Net model including a dynamic convolution kernel selection mechanism, introducing the dynamic convolution kernel selection mechanism in the Input module, and selecting the most suitable convolution kernel size according to different input features.

[0015] Furthermore, constructing an initial classification model based on the ResNet-18 algorithm includes designing a ResNet-18 model including multi-scale feature fusion, introducing a multi-scale feature extraction module in each residual block, and capturing images of different scales based on the multi-scale feature extraction module;

[0016] Convolution kernels of different sizes are introduced into the multi-scale feature extraction module and fused in the residual block.

[0017] Furthermore, the training of the initial classification model based on the third stored data to obtain classification models corresponding to different classification algorithms specifically includes:

[0018] After constructing different initial classification models based on different classification algorithms, selecting part of the data in the third storage data to train the initial classification model;

[0019] Calculate the indicators of each initial classification model, including accuracy, recall, and F1-score;

[0020] The model is evaluated according to the calculation results of each initial classification model indicator, and the performance of different SMT image data features in each initial classification model is analyzed based on the evaluation results. The feature grouping is associated with the initial classification model performance based on the analysis results.

[0021] Furthermore, the training of the initial classification model based on the third stored data to obtain classification models corresponding to different classification algorithms specifically includes:

[0022] According to the correlation result between the feature grouping and the initial classification model performance, the training data set of each initial classification model is determined based on the third stored data, and each initial classification model is further trained by the determined training data set to obtain classification models corresponding to different classification algorithms.

[0023] Furthermore, the step of selecting a target classification model from different classification models according to the SMT image data to be classified specifically includes:

[0024] A feature analysis is performed on each image of the SMT image data that needs to be classified, and a suitable classification model is selected from different classification models as the target classification model according to the result of the feature analysis.

[0025] Furthermore, after the target classification model is determined, feature extraction is performed on the SMT image data to be classified based on the characteristics of the target classification model to obtain an input data set.

[0026] Furthermore, the method further comprises:

[0027] When more than one target classification model is determined, each image to be classified is input into the target classification model, the prediction confidence of each target classification model is calculated, and the output result of the final classification model is selected as the final output result according to the prediction confidence.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The present invention obtains classification models by training a variety of classification algorithms respectively, and performs feature analysis on the image data to be classified, and matches a suitable classification model based on the characteristics of different classification models, thereby achieving efficient and automatic screening of SMT images and improving the accuracy of screening results. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings:

[0031] Figure 1 The present invention is a flowchart of an image data screening method based on multiple artificial intelligence classification algorithms disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments.

[0033] Embodiment 1

[0034] See also Figure 1 , Figure 1 : is a flow chart of an image data screening method based on multiple artificial intelligence classification algorithms disclosed in an embodiment of the present invention, the method comprising the following steps:

[0035] The SMT image data is collected, and the collected SMT image data is stored in a database as first storage data.

[0036] The first storage data is saved in a structure where the first-level folder is the component type, the second-level folder is the component subdivision model, and the third-level folder is the defect type; wherein the second-level folder is a subfolder of the first-level folder, and the third-level folder is a subfolder of the second-level folder.

[0037] A data processing operation is performed on the first stored data, and the processed data is stored as the second stored data in a structure where the first-level folder is the component type, the second-level folder is the defect type, and the third-level folder is the component subdivision model.

[0038] Specifically, in the embodiment of the present invention, the component type may be a resistor, a capacitor, an IC, etc., the subdivision model of the component may be a specific resistance value of the resistor such as 1kΩ, 10kΩ, etc., and the defect type includes but is not limited to welding defects, surface defects, etc. Before the data processing operation, storing the data in a first-level folder as the component type, a second-level folder as the subdivision model of the component, and a third-level folder as the defect type is helpful for systematically storing and managing the original data.

[0039] After data processing, the data is reorganized by defect type, that is, the first-level folder is organized by component type, the second-level folder is organized by defect type, and the third-level folder is organized by component subdivision model, so that defect classification training can be more convenient. In this way, the classification algorithm can more easily access different component data of the same type of defects, thereby improving the accuracy and generalization ability of the model. Among them, the data processing operations here include but are not limited to data cleaning, image enhancement, and defect annotation of the first stored data to remove noise, blurry or incomplete images, correct defects in the image, and enhance the image quality by adjusting brightness, contrast, and color balance, etc., to ensure data consistency and quality.

[0040] A primary screening operation is performed on the second storage data, wherein the primary screening operation includes screening the second storage data according to useful data and useless data, and the data after the primary screening is separately saved in a useful data folder and a useless data folder as the third storage data.

[0041] Specifically, in the embodiment of the present invention, useful data refers to image data with clear images, accurate defect annotations, no severe noise or damage, which can provide effective training information for the model. Useless data refers to image data with blurred images, incorrect defect annotations, severe noise or damage, which may affect the training effect of the model.

[0042] Image clarity detection algorithms, such as Laplace transform, are used to calculate the clarity value of the image. Images below a certain threshold are marked as useless data. Image noise detection algorithms are used to detect the noise level in the image. Images with noise levels above a certain threshold are marked as useless data. In addition, image integrity checks are also performed to detect the integrity of the image, such as whether there are defects or occlusions. Images with serious defects are marked as useless data. Furthermore, the accuracy of annotations is also checked to check the accuracy and consistency of defect annotations. Images with incorrect or incomplete annotations are marked as useless data.

[0043] After the automated screening operation, optionally, the useful data and useless data screened out by the automated screening tool are manually reviewed to ensure the accuracy of the screening results. Based on the review results, the screening criteria and parameters of the automated screening tool are adjusted to improve the accuracy of the automated screening.

[0044] Initial classification models are constructed respectively through different classification algorithms, and the initial classification models are trained based on the third storage data to obtain classification models corresponding to different classification algorithms. The classification algorithms include AlexNet algorithm, VGGNet-16 algorithm, GoogleNet algorithm and ResNet-18 algorithm;

[0045] A target classification model is selected from different classification models according to the SMT image data to be classified, and the screening result is output through the target classification model.

[0046] Furthermore, constructing an initial classification model based on the AlexNet algorithm includes designing an AlexNet model including local perception field enhancement, introducing multi-scale convolution kernels in some convolution layers, and increasing the size of the convolution kernels.

[0047] Specifically, some convolutional layers of AlexNet are selected for modification, and multi-scale convolution kernels are introduced and the size of the convolution kernels is increased. In the first convolutional layer of AlexNet, multiple convolution kernels of different sizes are introduced, such as 3x3, 5x5 and 7x7. These convolution kernels can be applied to the input image data at the same time, and their output results are fused. At this time, for each input image, convolution operations are performed using 3x3, 5x5 and 7x7 convolution kernels respectively to obtain three sets of feature maps, which are then concatenated to form a new feature map. In the second convolutional layer of AlexNet, the size of the convolution kernel is increased from 3x3 to 5x5 or 7x7, and the original 3x3 convolution kernel is replaced with a 5x5 or 7x7 convolution kernel to perform a convolution operation to obtain a feature map with a larger receptive field.

[0048] For example, the input image pixel size is 224x224. After convolution with a 3x3 convolution kernel, a 224x224 feature map is obtained. After convolution with a 5x5 convolution kernel, a 224x224 feature map is obtained. After convolution with a 7x7 convolution kernel, a 224x224 feature map is obtained. These three feature maps are concatenated in the channel dimension to obtain a feature map containing three scale information. The feature map output by the first convolution layer is input into the second convolution layer, and a 5x5 or 7x7 convolution kernel is used for convolution operation to obtain a feature map with a larger receptive field.

[0049] By introducing multi-scale convolution kernels and increasing the size of the convolution kernel in AlexNet, the model's feature extraction capability can be enhanced and the receptive field can be expanded, thereby better adapting to the characteristics of SMT image data and improving the performance of the classification model.

[0050] Furthermore, constructing an initial classification model based on the VGGNet-16 algorithm includes designing a VGGNet-16 model including deep feature fusion, introducing skip connections in the middle layer of the network, and fusing shallow features with deep features.

[0051] VGGNet-16 is a classic deep convolutional neural network, consisting of multiple convolutional layers and pooling layers. The core is to use 3x3 convolution kernels and 2x2 pooling kernels to build a deep network structure by stacking multiple convolutional layers. Based on the original VGGNet-16, we introduced the design of deep feature fusion to enhance its classification ability. Introducing additional jump connections in the middle layer of VGGNet-16 can directly connect shallow feature maps with subsequent deeper feature maps. Specifically, after the convolutional layer of the network, that is, after the output of each convolutional layer, a jump connection is introduced to connect the output of the current layer with the output of some previous layers.

[0052] For example, assuming that a skip connection is introduced after the third convolutional block to fuse the output of the first convolutional block with the output of the third convolutional block, the shallow features represent low-level local image features, such as edges and textures, while the deep features represent more abstract and high-level semantic information, such as object parts and overall shapes.

[0053] Suppose you want to classify SMT images, which include different types of electronic components such as resistors, capacitors, and integrated circuits. During the training process, VGGNet-16 can better capture and fuse the feature representations of different levels of these different types of electronic components by introducing skip connections and deep feature fusion.

[0054] Skip connections and deep feature fusion enable the model to extract features from multiple levels, thereby improving the model's ability to understand complex image structures. For example, it is possible to capture both the detailed features of components (such as package type) and the overall structure (such as the arrangement on a circuit board) at the same time. In addition, by retaining and transferring shallow features, the model is able to more comprehensively consider local details in the image, thereby reducing information loss and improving classification accuracy and robustness.

[0055] Furthermore, building an initial classification model based on the Google Net algorithm includes designing a Google Net model including a dynamic convolution kernel selection mechanism, introducing the dynamic convolution kernel selection mechanism in the Input module, and selecting the most suitable convolution kernel size according to different input features.

[0056] In an embodiment of the present invention, the Interception module in GoogleNet is selected for modification, and a dynamic convolution kernel selection mechanism is introduced. A dynamic convolution kernel selection mechanism is introduced in the Interception module, which selects the most suitable convolution kernel size according to different input features. Specifically, it includes a plurality of convolution kernels of different sizes (for example, 1x1, 3x3, 5x5, 7x7), and a selector is used to select the convolution kernel that best suits the current input feature before the convolution operation. Specifically, based on the selector module, the current feature map is input, and the most suitable convolution kernel size is output through the selector module. The selector can be implemented by a lightweight neural network, and the weight of each convolution kernel size is calculated using the input feature map, and the most suitable convolution kernel size is selected according to the weight.

[0057] For example, the input feature map is SMT image data of size 28x28. In the Inception module, 1x1, 3x3, 5x5 and 7x7 convolution kernels are preset, and these convolution kernels are applied in parallel to perform convolution operations on the 28x28 feature map to obtain four sets of feature maps. Based on the 28x28 feature map input to the selector module, the weights of each convolution kernel size (1x1, 3x3, 5x5, 7x7) are calculated through a fully connected layer or convolution layer, such as [0.1, 0.5, 0.3, 0.1]. According to the weights output by the selector, the four sets of feature maps are weighted summed to obtain the final feature map. For example, the weight of the 1x1 convolution kernel output feature map is 0.1, the weight of the 3x3 convolution kernel output feature map is 0.5, the weight of the 5x5 convolution kernel output feature map is 0.3, and the weight of the 7x7 convolution kernel output feature map is 0.1. According to the weights, the most suitable convolution kernel size is 3x3 convolution kernel.

[0058] In the embodiment of the present invention, the most suitable convolution kernel size can be selected according to different input features through the dynamic convolution kernel selection mechanism, thereby enhancing the flexibility and accuracy of feature extraction. For example, for small defects, a smaller convolution kernel (such as 1x1 or 3x3) is selected for feature extraction; for larger defects, a larger convolution kernel (such as 5x5 or 7x7) is selected for feature extraction.

[0059] Further, as a preferred implementation of Example 1 of the present invention, the four groups of feature maps are weighted and summed to obtain the final feature map, i.e., the fused feature map, and then the fused feature map is subjected to a convolution operation using a dynamic convolution kernel selection mechanism. The dynamic convolution kernel selection mechanism can adaptively select the size of the convolution kernel according to the spatial features and semantic information of the feature map to maximize the effect of feature extraction.

[0060] By weighted summing and fusing feature maps of different scales and levels, the model's ability to represent the input image can be effectively enhanced. This multi-scale feature fusion helps capture richer and more abstract image features. In addition, the dynamic convolution kernel selection mechanism can dynamically adjust the size and shape of the convolution kernel according to the specific features of the input image, which can better adapt to SMT images of different sizes and complexities, thereby improving the generalization ability and classification accuracy of the model.

[0061] Furthermore, constructing an initial classification model based on the ResNet-18 algorithm includes designing a ResNet-18 model including multi-scale feature fusion, introducing a multi-scale feature extraction module in each residual block, and capturing images of different scales based on the multi-scale feature extraction module;

[0062] Convolution kernels of different sizes are introduced into the multi-scale feature extraction module and fused in the residual block.

[0063] In the embodiment of the present invention, by introducing a multi-scale feature extraction module, detailed information and semantic features at different scales can be captured, so that the model has stronger representation capabilities when processing complex scenes. And by effectively integrating multi-scale features, the ResNet-18 algorithm classification model can improve the accuracy of classification tasks and improve the adaptability to data of different scales and complexity.

[0064] It should be noted that although both the Google Net algorithm and the ResNet-18 algorithm involve the processing of multi-scale features and improving the model's ability to process complex data, Google Net focuses on dynamically selecting convolution kernels to adapt to the needs of different input features, while ResNet-18 focuses on introducing multi-scale feature extraction and fusion mechanisms in each residual block, each optimizing the model's feature extraction and processing capabilities in different directions to adapt to different image classifications.

[0065] Furthermore, the initial classification model is trained based on the third stored data to obtain classification models corresponding to different classification algorithms, specifically including:

[0066] After different initial classification models are constructed based on different classification algorithms, part of the data in the third storage data is selected to train the initial classification model, and indicators of each initial classification model are calculated, including accuracy, recall rate, and F1-score.

[0067] Specifically, first use part of the data in the third storage data set as the training set and validation set to train different initial classification models. After the training is completed, the performance of each model is evaluated, and the classification effect of each model on the validation set or test set is measured by indicators such as accuracy, recall, and F1-score. For each model, analyze its performance on different features, including how the model recognizes and distinguishes various features in SMT image data, such as defect types, component segmentation, etc. Compare and analyze the performance of different models on the same data set to find out which models perform better in specific feature recognition and which models may have limitations in certain features or need further optimization.

[0068] Furthermore, the model is evaluated according to the calculation results of each initial classification model indicator, and the performance of different SMT image data features in each initial classification model is analyzed according to the evaluation results, and the feature grouping is associated with the initial classification model performance based on the analysis results.

[0069] Specifically, various features in the SMT image data are classified and grouped according to the results of feature analysis. For example, different defect types, component subdivisions and other features can be divided into different groups. Each feature group is associated with its performance on different initial classification models, and a mapping relationship between features and model adaptability is established to ensure that each model can perform optimally when processing specific features.

[0070] Furthermore, training the initial classification model based on the third stored data to obtain the classification models corresponding to different classification algorithms specifically includes:

[0071] According to the correlation result between the feature grouping and the initial classification model performance, the training data set of each initial classification model is determined based on the third stored data, and each initial classification model is further trained by the determined training data set to obtain classification models corresponding to different classification algorithms.

[0072] Specifically, different features in the SMT image data are grouped, such as different types of defects, components of different subdivisions, etc. Based on the performance of the previous model, the adaptability and advantages of each classification model on different feature groups are determined. Based on the feature grouping results, the most suitable training data set is selected for each initial classification model. These training data sets contain SMT image data of specific feature groups to ensure that the model can learn and optimize the classification capabilities for these features during the training process.

[0073] By selecting and optimizing the training data set in a targeted manner, the classification accuracy and generalization ability of each classification model on a specific feature group can be significantly improved. At the same time, by accurately selecting the training data set, it helps to optimize the utilization efficiency of computing resources and reduce unnecessary training time and costs.

[0074] As a preferred implementation of Example 1 of the present invention, the classification models corresponding to different algorithms are specifically as follows:

[0075]

[0076] Where x is the input SMT image data; Conv3(x), Conv5(x), and Conv7(x) represent convolution operations using 3x3, 5x5, and 7x7 convolution kernels, respectively; is a feature concatenation operation, which concatenates the outputs of different convolution kernels; b1 is the bias term after the convolution operation; ReLU(·) is the activation function; MaxPool(·) is the maximum pooling operation; W1 and W2 are the weight matrices of the subsequent fully connected layers; b2 and b3 are the bias terms of the subsequent fully connected layers; f Alexnet (x) is the output of the AlexNet model.

[0077] f VGGNet-16 (x)=W3·(ReLU(W4·(MaxPool(ReLU(W5·x+b4)))+

[0078] SkipConnection(w5·x)+b5)))+b6

[0079] Where x is the input SMT image data; W5·x+b4 is the convolution operation, w5 is the convolution kernel weight, b4 is the bias term; ReLU(·) is the activation function; MaxPool(·) is the maximum pooling operation;

[0080] SkipConnection (w5 x) is a skip connection that fuses shallow features with deep features; W3 and W4 are weight matrices of subsequent fully connected layers; b5 and b6 are bias items of subsequent fully connected layers; f VGGNet-16 (x) is the output of the VGGNet-16 model.

[0081] f GoogleNet (x)=W6·(ReLU(W7·(DynamicConv(x)+b7)))+b8

[0082]

[0083] Among them, x is the input SMT image data; DynamicConv(x) is a dynamic convolution operation, which selects the most suitable convolution kernel size according to the input features; α i To select weights; Conv ki (x) is the convolution operation with different convolution kernel sizes; ReLU(·) is the activation function; W6 and W7 are the weight matrices of the subsequent fully connected layers; b7 and b8 are the bias items of the subsequent fully connected layers; fGoogleNet (x) is the output of the GoogleNet model.

[0084]

[0085] Block(x)=ReLU(x+MuitiScale(W9·ReLU(W 10 ·x+b 10 ))+b 11 )

[0086] Where x is the input SMT image data; Block is the multi-scale residual block; MuitiScale is a multi-scale feature extraction module, which is used to extract features in convolution operations with different convolution kernel sizes; W 10 ·x+b 10 is the convolution operation, W 10 is the convolution kernel weight, b 10 is the bias term; ReLU(·) is the activation function; W8 and W9 are the weight matrices of the subsequent fully connected layers; b9 and b 11 is the bias term of the subsequent fully connected layer; ° is a function composite operation, indicating the superposition of multi-scale residual blocks; f ResNet-18 (x) is the output of the ResNet-18 model.

[0087] Furthermore, the step of selecting a target classification model from different classification models according to the SMT image data to be classified specifically includes:

[0088] A feature analysis is performed on each image of the SMT image data that needs to be classified, and a suitable classification model is selected from different classification models as the target classification model according to the result of the feature analysis.

[0089] Specifically, according to the results of feature analysis, the target classification model that best suits the current image features is selected from different pre-built initial classification models, and the selection operation is performed based on the feature similarity measure, the performance evaluation of the model, and the model performance in the previous training process.

[0090] Furthermore, after determining the target classification model, feature extraction is performed on the SMT image data to be classified based on the characteristics of the target classification model to obtain an input data set. The purpose of this step is to ensure that the data input to the model meets its expected input format and requirements.

[0091] Furthermore, the method further comprises:

[0092] When more than one target classification model is determined, each image to be classified is input into the target classification model, the prediction confidence of each target classification model is calculated, and the output result of the final classification model is selected as the final output result according to the prediction confidence.

[0093] Finally, it should be noted that the image data screening method based on multiple artificial intelligence classification algorithms disclosed in the embodiment of the present invention only discloses a preferred embodiment of the present invention, which is only used to illustrate the technical solution of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments can still be modified, or some of the technical features therein can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for screening image data based on multiple artificial intelligence classification algorithms, characterized in that: The method comprises the following steps: Collecting SMT image data, and storing the collected SMT image data in a database as first storage data; The first storage data is saved in a structure where the first-level folder is the component type, the second-level folder is the component subdivision model, and the third-level folder is the defect type; wherein the second-level folder is a subfolder of the first-level folder, and the third-level folder is a subfolder of the second-level folder; Performing data processing operations on the first stored data, and storing the processed data in a structure where the first-level folder is the component type, the second-level folder is the defect type, and the third-level folder is the component subdivision model as the second stored data; Performing a primary screening operation on the second stored data, the primary screening operation comprising screening the second stored data according to useful data and useless data, and saving the data after the primary screening separately in a useful data folder and a useless data folder as the third stored data; Initial classification models are constructed respectively by different classification algorithms, and part of the data in the third storage data is selected to train the initial classification models; the indicators of each initial classification model are calculated, and the indicators include accuracy, recall rate, and F1-score; the model is evaluated according to the calculation results of each initial classification model indicator, and the performance of different SMT image data features in each initial classification model is analyzed according to the evaluation results, and the feature grouping is associated with the initial classification model performance based on the analysis results; And according to the correlation result of the feature grouping and the initial classification model performance, the training data set of each initial classification model is determined based on the third storage data, and each initial classification model is further trained by the determined training data set to obtain classification models corresponding to different classification algorithms; the classification algorithms include AlexNet algorithm, VGGNet-16 algorithm, GoogleNet algorithm and ResNet-18 algorithm; Perform feature analysis on each image of the SMT image data that needs to be classified, select a suitable classification model from different classification models as the target classification model based on the result of the feature analysis, and output the screening result through the target classification model.

2. The image data screening method based on multiple artificial intelligence classification algorithms according to claim 1 is characterized in that: Building an initial classification model based on the AlexNet algorithm includes designing an AlexNet model with local perception field enhancement, introducing multi-scale convolution kernels in some convolution layers, and increasing the size of the convolution kernels.

3. The image data screening method based on multiple artificial intelligence classification algorithms according to claim 1 is characterized in that: Building the initial classification model based on the VGGNet-16 algorithm includes designing a VGGNet-16 model that includes deep feature fusion, introducing jump connections in the middle layer of the network, and fusing shallow features with deep features.

4. The image data screening method based on multiple artificial intelligence classification algorithms according to claim 1 is characterized in that: Building the initial classification model based on the GoogleNet algorithm includes designing a GoogleNet model with a dynamic convolution kernel selection mechanism, introducing a dynamic convolution kernel selection mechanism in the Inception module, and selecting the most suitable convolution kernel size according to different input features.

5. The image data screening method based on multiple artificial intelligence classification algorithms according to claim 1 is characterized in that: Building an initial classification model based on the ResNet-18 algorithm includes designing a ResNet-18 model including multi-scale feature fusion, introducing a multi-scale feature extraction module in each residual block, and capturing images of different scales based on the multi-scale feature extraction module; Convolution kernels of different sizes are introduced into the multi-scale feature extraction module and fused in the residual block.

6. The image data screening method based on multiple artificial intelligence classification algorithms according to any one of claims 1 to 5, characterized in that: After determining the target classification model, feature extraction is performed on the SMT image data to be classified based on the characteristics of the target classification model to obtain the input data set.

7. The image data screening method based on multiple artificial intelligence classification algorithms according to claim 6 is characterized in that: The method further comprises: When more than one target classification model is determined, each image to be classified is input into the target classification model, the prediction confidence of each target classification model is calculated, and the output result of the final classification model is selected as the final output result according to the prediction confidence.

Images