Deep multi-scale fusion-based chrysanthemum image intelligent screening method under meta-learning framework

Abstract

The invention discloses a chrysanthemum image intelligent screening method based on deep multi-scale fusion under a meta-learning framework, and the method comprises the steps: obtaining deep fusion features through a deep feature extraction module based on dense connection; taking the deep fusion features as feature input of a semantic analysis module to obtain semantic features; a deep feature extraction module, a semantic analysis module and a linear connection layer jointly form a backbone network, a parameter updating mode of the backbone network is learned and simulated through a meta-learning framework, and obtained priori knowledge is stored; and inputting the chrysanthemum distorted image into the backbone network, fusing the chrysanthemum distorted image into the backbone network through priori knowledge, and then performing fine tuning to obtain a model suitable for chrysanthemum quality evaluation. According to the method, prior knowledge can be effectively obtained from different types of distorted images and fused into a chrysanthemum quality evaluation task, meanwhile, the chrysanthemum quality can be accurately predicted by using a small amount of label data, so that quality screening is carried out, and the method has the advantages of high efficiency, robustness and high availability.

Description

Technical Field

[0001] The present invention belongs to the field of image quality evaluation, and specifically relates to an intelligent screening method for chrysanthemum images based on deep multi-scale fusion under a meta-learning framework. Background Art

[0002] With the development of computer technology and artificial intelligence, image quality assessment technology has become an important research direction in the field of computer vision. A good image quality assessment method can simulate human subjective habits to score the target image, and can quickly and accurately process large quantities of image data, which greatly reduces the cost of manually evaluating the quality of a large number of images and has high application value in the field of computer vision. Image quality assessment methods based on deep neural networks have quickly become popular with their good accuracy. Many mainstream image quality assessment methods have achieved remarkable results, which also confirms the effectiveness of deep neural networks. These methods show an important point in the learning process: the semantic analysis stage in the model learning process greatly affects the evaluation effect. Many image quality assessment methods improve the accuracy of the semantic analysis stage through feature fusion, feature interaction, network branching and other schemes. With the rise of Transformer, the semantic analysis stage in visual tasks has been promoted to a new level, and many improved image quality assessment methods have also emerged. These methods have indeed achieved good results. However, whether it is a traditional neural network or a Transformer network, the training process requires a large amount of labeled data, and the expected convergence model depends on the cost of data investment. How to more efficiently utilize effective knowledge from existing tasks to reduce dependence on large amounts of data is very important.

[0003] Meta-learning is a branch of machine learning, and its main goal is to make the machine learning algorithm itself a learner that can quickly adapt to new tasks or new environments. Instead of solving specific tasks directly, it focuses on how to optimize the learning process itself through empirical learning, so as to learn how the algorithm adjusts parameters or strategies when solving new tasks to quickly obtain good performance. This ability to "learn how to learn" enables the algorithm to learn efficiently with a small amount of sample data. Images with different distortion types and distorted images of chrysanthemums have commonalities in the laws of quality fluctuation. If meta-learning can be used to obtain effective prior knowledge from existing data and reasonably integrated into the task of chrysanthemum quality evaluation, it may be possible to reduce dependence on chrysanthemum quality labels while obtaining better prediction performance. Summary of the invention

[0004] Purpose of the invention: In order to overcome the deficiencies of the prior art, a method for intelligent screening of chrysanthemum images based on deep multi-scale fusion under a meta-learning framework is provided. In order to realize the semantic analysis of multi-region features by the network and improve the accuracy of the model in evaluating distorted images under quality fluctuations, a semantic analysis network based on the self-attention mechanism is designed, a random recombination strategy is formulated to adjust the depth feature size, reduce the parameter burden brought by the self-attention mechanism, and provide data support for the network in the semantic analysis stage. At the same time, in order to adapt to the needs of real tasks and improve the generalization of the model in multi-type quality evaluation tasks, a cross-image type quality model parameter optimization scheme is designed under the meta-learning framework, which can accurately evaluate the quality of multi-type images and has the ability to evaluate image quality across image types. It has the advantages of being objective, fast, and easy to use, and provides a scientific basis for multi-type image quality evaluation methods.

[0005] Technical solution: To achieve the above purpose, the present invention provides a method for intelligent screening of chrysanthemum images based on deep multi-scale fusion under a meta-learning framework, comprising the following steps:

[0006] S1: Through the constructed densely connected deep feature extraction module, three different levels of features of the distorted image are extracted, and the three different levels of features are spliced ​​as deep fusion features under different receptive fields;

[0007] S2: Use the deep fusion features as the feature input of the constructed semantic analysis module based on Vision Transformer, analyze the mutual relationship and semantic information between multi-level deep features, and obtain semantic features;

[0008] S3: The semantic features are passed through a linear connection layer to obtain a quality score. The loss function of the backbone network is the mean square error loss, and the Adam optimizer is used for gradient calculation and back propagation.

[0009] S4: The deep feature extraction module, the semantic analysis module and the linear connection layer are combined to form a backbone network. During the execution of steps S1 to S3, the constructed meta-learning framework is used to learn and simulate the parameter update mode of the backbone network. The distorted image dataset is used as a meta-training object, and the experience of evaluating the image quality rules is summarized from the distorted images. The meta-learning framework periodically learns each meta-task according to the learning step size and saves the learned prior knowledge.

[0010] S5: Input the distorted image of chrysanthemum to be identified into the backbone network, and fine-tune it after integrating prior knowledge into the backbone network to obtain a model suitable for chrysanthemum quality evaluation, obtain the chrysanthemum quality evaluation result, and perform quality screening on the chrysanthemum image.

[0011] Furthermore, the deep feature extraction module in step S1 is composed of head feature extraction, 3 dense modules and 3 transition layers. Among them, the head feature extraction is composed of a convolution layer, a normalization layer, an activation function and a maximum pooling layer, which is used to preliminarily adjust the feature size and reduce the feature size to 56*56. The three dense modules and the three transition layers alternately complete the feature extraction task, wherein the three dense modules are respectively composed of 6, 12, and 24 dense layers, each dense layer is composed of two groups of convolution layers, normalization layers, and activation functions, which meet the extraction of the current receptive field features, and the input and output features of each dense layer are spliced ​​with channel features as the input of the next dense layer, so as to reduce the gradient disappearance and increase the richness of feature transfer. The transition layer after each dense module is composed of a convolution layer, a normalization layer, an activation function and average pooling, which is used to reduce the size of the overall feature and the number of channels. The transition layers at three different levels will obtain features of 56*56, 28*28, and 14*14 sizes respectively. The 56*56 and 28*28 features are convolved to obtain 14*14 features. The 14*14 features at three different levels are channel-joined to obtain the initial fused deep features. The deep features contain feature information at different levels, reducing the gradient transfer burden caused by long-distance dependencies, and providing the overall network with richer receptive field features and sources of semantic analysis.

[0012] Furthermore, in step S2, a semantic analysis module is designed, which uses PositionEmbedding to annotate the position information of different channel features for subsequent analysis of the weight relationship between different features. After reducing the two-dimensional features to one dimension, a serialized input is formed, and the feature semantics is analyzed using Encoder. The Multi-headAttention in the Encoder assigns Q, K, and V features to each sequence feature, and the weight mapping relationship between different features is obtained by weighting. On the basis of the attention mechanism, each position in the input sequence is allowed to pay attention to other positions in the sequence, and the weighted average is dynamically calculated as the output, so as to capture the dependency relationship in the sequence, thereby further obtaining the overall and local contextual relationship of different channel features, and finally obtaining the semantic information and weight relationship between different features. The process of obtaining the weight relationship feature by Multi-headAttention is as follows, which is the same as normalizing the obtained weights to represent the dimension size:

[0013]

[0014] Furthermore, in step S3, in order to better obtain the semantic features obtained by S2 and obtain the corresponding relationship between image quality fluctuations and scores, a linear connection layer is designed. The linear connection layer includes multiple channel connection layers to meet the mapping requirements of the overall channel from 1000 to 512, 512 to 256, 256 to 32, and 32 to 1. Normalization operations and activation functions are interspersed between channel mappings to reduce the extreme values ​​of the values, which is conducive to more reasonable numerical analysis and can obtain more accurate quality scores. In order to scientifically express the loss between the predicted value and the true value during the training process, and to better reflect the convergence process of the model, the mean square error loss is selected as the loss calculation. At the same time, the Adam optimizer is designed to calculate the overall gradient change and perform back propagation to gradually obtain a model with better performance until convergence. Among them, the loss function is calculated as follows, where is the quality score predicted by the model and is the true quality score corresponding to the image:

[0015] Loss=||y pre -y|| 2

[0016] Furthermore, in step S4, the meta-learning framework updates parameters by periodically learning the network. The specific process is as follows: the images with the same distortion type in the distorted images are classified into the same meta-task T i , the images in the meta-task are divided into the support set in a 4:1 ratio and queryset When the model is trained on both the support set and the query set, meta-learning will learn and summarize the prior knowledge. i The support set in and queryset Divide into multiple subtasks and The rule of meta-learning summarizing prior knowledge is as follows: each time a single meta-task is trained, the framework analyzes some network parameters according to the set learning step size. When a set number of tasks or the current training stage is completed, meta-learning updates the accumulated prior knowledge to the parameter dictionary. This process not only helps meta-learning learn from different distorted information, but also provides key information units for forming prior knowledge. When all training tasks are completed, the meta-learning framework completes the learning of the network model and summarizes the regional connections and semantic information of all types of images, forming prior knowledge with generalization and stability. Different learning rates ∈ 2 are applied to the two module parts and the linear connection layer during the meta-training process. fnet ,∈ Linear For network update, when the network trains the jth subtask, the parameter update process of the multi-channel feature extraction network and the inter-layer fusion network is:

[0017]

[0018]

[0019] in, and They represent the parameter update results of the j-th support set subtask and the j-1-th support set subtask in the i-th meta-task respectively; the role of ε is to avoid the divisor being 0; and They represent the first-order moment estimate and the second-order moment estimate after bias correction in the j-th task, respectively. and The specific calculation process is as follows:

[0020]

[0021]

[0022] The parameter update process of the linear regression part is the same as above, just change ∈ fnet The corresponding learning rate is replaced by ∈ Linear The corresponding parameters can be obtained; the learning method of meta-learning is controlled within the step size σ of the cycle, and the memory capacity of meta-learning will also change with the sequence j of each subtask; when training each subtask, from each meta-task T i The prior knowledge θ learned in i The change is defined as:

[0023]

[0024] When all meta-tasks T i After the training and learning of are completed, the prior knowledge θ containing all the distortion information is formed. prior .

[0025] Furthermore, in step S5, the prior knowledge obtained by the meta-learning framework in the distorted image is used in the subsequent fine-tuning. In order to improve the accuracy of chrysanthemum quality prediction and reduce the dependence on chrysanthemum data labels, the prior knowledge obtained by meta-learning is imported into the backbone network, and the backbone network is retrained to quickly fine-tune the model used for chrysanthemum quality evaluation. In view of the high efficiency and accuracy of the fine-tuning process, ∈ fnet and ∈ Linear Set to the same value ∈, that is, the module part of the backbone network and the linear connection layer use the same learning rate. During the fine-tuning process, the process of the model obtaining the chrysanthemum quality score is:

[0026] y pre =f(θ prior ,∈)

[0027] Meta-learning is a branch of machine learning, and its main goal is to make the machine learning algorithm itself a learner that can quickly adapt to new tasks or new environments. It does not directly solve specific tasks, but focuses on how to optimize the learning process itself through empirical learning. Meta-learning focuses on how the learning algorithm adjusts parameters or strategies when solving new tasks to quickly obtain good performance. This ability to "learn how to learn" enables the algorithm to learn efficiently with a small amount of sample data, providing new solutions for dealing with complex tasks and problems. It is a promising and dynamic research direction in the field of artificial intelligence.

[0028] The present invention utilizes the dense connections in the deep feature extraction module to accurately extract features, and the frequent short connections enrich the feature information. At the same time, a transition layer is added to reduce the feature dimension. Finally, the features of three different levels are spliced ​​as deep fusion features under different receptive fields.

[0029] The present invention uses a semantic analysis network to analyze the weight relationship between different features, and captures the dependencies in the sequence based on the attention mechanism, so as to further obtain the overall and local contextual relationships of different channel features.

[0030] The meta-learning framework designed in the present invention enables the model to quickly learn new tasks with a small amount of sample data. By learning learning strategies that adapt to different tasks, it can achieve better performance when seeing very small amounts of data, and rely less on manual adjustment or hyperparameter selection, thereby reducing the need for human intervention and improving the degree of automation and intelligence.

[0031] The overall model and algorithm framework of the present invention can effectively utilize the prior knowledge of different distorted images to complete the evaluation of chrysanthemum quality, achieve ideal evaluation results in different amounts of labeled data, and meet the actual needs of chrysanthemum quality evaluation and identification in reality.

[0032] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0033] 1. Compared with the subjective chrysanthemum quality evaluation method which is time-consuming, unstable and relies on manual experience, the present invention has the advantages of high efficiency and real-time.

[0034] 2. Compared with the existing method of selecting a single network for feature extraction, the present invention considers the relationship between feature information of different channels, explores the semantic characteristics of different features, and can obtain more accurate evaluation results.

[0035] 3. Compared with the existing traditional network models, the present invention separates the shallow feature extraction part from the semantic analysis part, and obtains modules suitable for deep feature extraction and semantic analysis respectively, which can realize efficient learning of the two processes and improve the accuracy of model analysis.

[0036] 4. Compared with traditional training with a single type of data, the present invention also focuses on mining effective data analysis rules from data sets of different types that can still provide relevant knowledge and information. It makes use of the common relationships between different types of images to a large extent, realizes efficient use of data, and satisfies the compatibility between different knowledge.

[0037] 5. Compared with the existing image quality assessment methods across image types, the present invention fully integrates the prior knowledge obtained by meta-learning when training the network across different image types, requiring only a small amount of time for fine-tuning and less labeled data for training, thus reducing dependence on labels. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flow chart of the method of the present invention;

[0039] Figure 2 The results of predicting chrysanthemums of different quality levels according to the embodiment of the present invention, wherein: (a), (b), (c), (d), and (e) correspond to chrysanthemum images from high to low quality, respectively;

[0040] Figure 3 The scatter plots of correlation between predicted values ​​and true values ​​of the embodiments of the present invention, wherein: (a) is the correlation scatter plot without multi-level feature fusion; (b) is the correlation scatter plot obtained by the complete algorithm processing of the present invention; the sample points circled in red in the figure are outliers. DETAILED DESCRIPTION

[0041] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0042] The present invention provides a chrysanthemum image intelligent screening method based on deep multi-scale fusion under a meta-learning framework. Figure 1 As shown, it includes the following steps:

[0043] S1: Design and construct a deep feature extraction module based on dense connections, use frequent feature short connections to enrich feature information, and add transition layers to reduce feature dimensions; extract three different levels of features from the distorted image through the constructed deep feature extraction module based on dense connections, and splice the three different levels of features as deep fusion features under different receptive fields;

[0044] S2: Design and build a semantic analysis module based on Vision Transformer, and use the obtained deep fusion features as the feature input of the semantic analysis module. Since step S1 obtains the channel features after block division, the Patch Embedding part of Vision Transformer is removed; Position Embedding is performed on each channel feature of the deep fusion features to mark the position information of different channel features, so as to update the weight relationship between features later, and the semantic features are obtained after being processed by 12 Encoders;

[0045] S3: Design and construct a linear connection layer. After the semantic features pass through the linear connection layer, the quality score can be obtained. The deep feature extraction module, semantic analysis module and linear connection layer together form a backbone network. The loss function of the backbone network is the mean square error loss. The Adam optimizer is used for gradient calculation and back propagation.

[0046] S4: Design a meta-learning framework to learn and simulate the network's parameter update mode, use the distorted image dataset as the meta-training object, summarize the experience of evaluating image quality rules from the distorted images, and periodically learn each meta-task according to the learning step size, and save the learned prior knowledge;

[0047] S5: Input the distorted chrysanthemum image to be evaluated into the backbone network, integrate the obtained prior knowledge into the backbone network and then perform fine-tuning to obtain a model suitable for chrysanthemum quality evaluation, obtain the chrysanthemum quality evaluation result, and perform quality screening on the chrysanthemum image.

[0048] In step S1:

[0049] The deep feature extraction module consists of head feature extraction, 3 dense modules and 3 transition layers. Among them, the head feature extraction consists of convolution layer, normalization layer, activation function and maximum pooling layer, which is used to preliminarily adjust the feature size and reduce the feature size to 56*56. The 3 dense modules and 3 transition layers alternately complete the feature extraction task. The 3 dense modules are composed of 6, 12 and 24 dense layers respectively. Each dense layer is composed of two groups of convolution layers, normalization layers and activation functions to meet the extraction of the current receptive field features. The input and output features of each dense layer are spliced ​​with channel features as the input of the next dense layer, so as to reduce the gradient disappearance and increase the richness of feature transfer. The transition layer after each dense module is composed of convolution layer, normalization layer, activation function and average pooling, which is used to reduce the size and number of channels of the overall features. The transition layers at three different levels will obtain features of 56*56, 28*28, and 14*14 sizes respectively. The 56*56 and 28*28 features are convolved to obtain 14*14 features. The 14*14 features at three different levels are channel-joined to obtain the initial fused deep features. The deep features contain feature information at different levels, reducing the gradient transfer burden caused by long-distance dependencies, and providing the overall network with richer receptive field features and sources of semantic analysis.

[0050] In step S2:

[0051] The semantic analysis module uses PositionEmbedding to annotate the position information of different channel features for subsequent analysis of the weight relationship between different features. After reducing the two-dimensional features to one dimension, a serialized input is formed, and the feature semantics is analyzed using Encoder. Multi-headAttention in Encoder assigns Q, K, and V features to each sequence feature, and the weight mapping relationship between different features is obtained by weighting. Based on the attention mechanism, each position in the input sequence is allowed to pay attention to other positions in the sequence, and the weighted average is dynamically calculated as the output to capture the dependency relationship in the sequence, thereby further obtaining the overall and local contextual relationship of different channel features, and finally obtaining the semantic information and weight relationship between different features. The process of obtaining the weight relationship feature by Multi-headAttention is as follows, which is the same as normalizing the obtained weights to represent the dimension size:

[0052]

[0053] In step S3:

[0054] In order to better obtain the semantic features through S2 and obtain the corresponding relationship between image quality fluctuations and scores, a linear connection layer is designed. The linear connection layer contains multiple channel connection layers to meet the mapping requirements of the overall channel from 1000 to 512, 512 to 256, 256 to 32, and 32 to 1. Normalization operations and activation functions are interspersed between channel mappings to reduce the extreme values, which is conducive to more reasonable numerical analysis and more accurate quality scores. In order to scientifically express the loss between the predicted value and the true value during the training process, and to better reflect the convergence process of the model, the mean square error loss is selected as the loss calculation. At the same time, the Adam optimizer is designed to calculate the overall gradient change and perform backpropagation to gradually obtain a model with better performance until convergence. Among them, the loss function is calculated as follows, y pre is the quality score predicted by the model, and y is the actual quality score corresponding to the image:

[0055] Loss=||y pre -y|| 2

[0056] In step S4:

[0057] The meta-learning framework updates parameters of the learning network periodically. The specific process is as follows: the images with the same distortion type in the distorted images are classified into the same meta-task T i , the images in the meta-task are divided into the support set in a 4:1 ratio and queryset When the model is trained on both the support set and the query set, meta-learning will learn and summarize the prior knowledge. i The support set in and queryset Divide into multiple subtasks and The rule of meta-learning summarizing prior knowledge is as follows: each time a single meta-task is trained, the framework analyzes some network parameters according to the set learning step size. When a set number of tasks or the current training stage is completed, meta-learning updates the accumulated prior knowledge to the parameter dictionary. This process not only helps meta-learning learn from different distorted information, but also provides key information units for forming prior knowledge. When all training tasks are completed, the meta-learning framework completes the learning of the network model and summarizes the regional connections and semantic information of all types of images, forming prior knowledge with generalization and stability. Different learning rates ∈ 2 are applied to the two module parts and the linear connection layer during the meta-training process. fnet ,∈ Linear For network update, when the network trains the jth subtask, the parameter update process of the multi-channel feature extraction network and the inter-layer fusion network is:

[0058]

[0059]

[0060] in, and They represent the parameter update results of the j-th support set subtask and the j-1-th support set subtask in the i-th meta-task respectively; the role of ε is to avoid the divisor being 0; and They represent the first-order moment estimate and the second-order moment estimate after bias correction in the j-th task, respectively. and The specific calculation process is as follows:

[0061]

[0062]

[0063] The parameter update process of the linear regression part is the same as above, just change ∈ fnet The corresponding learning rate is replaced by ∈ Linear The corresponding parameters can be obtained; the learning method of meta-learning is controlled within the step size σ of the cycle, and the memory capacity of meta-learning will also change with the sequence j of each subtask; when training each subtask, from each meta-task T i The prior knowledge θ learned in i The change is defined as:

[0064]

[0065] When all meta-tasks T i After the training and learning of are completed, the prior knowledge θ containing all the distortion information is formed. prior .

[0066] In step S5:

[0067] The prior knowledge obtained by the meta-learning framework in the distorted image is used in subsequent fine-tuning. In order to improve the accuracy of chrysanthemum quality prediction and reduce the dependence on chrysanthemum data labels, the prior knowledge θ obtained by the meta-learning framework in the distorted image is used in subsequent fine-tuning. In order to improve the accuracy of chrysanthemum quality prediction and reduce the dependence on chrysanthemum data labels, the prior knowledge θ obtained by the meta-learning framework is used in subsequent fine-tuning. prior Import it into the backbone network, and use the backbone network retraining process to quickly fine-tune the model used for chrysanthemum quality evaluation. In view of the efficiency and accuracy of the fine-tuning process, ∈ fnet and ∈ LinearSet to the same value ∈, that is, the module part of the backbone network and the linear connection layer use the same learning rate. During the fine-tuning process, the process of the model obtaining the chrysanthemum quality score is:

[0068] y pre =f(θ prior ,∈)

[0069] Based on the above steps, in order to verify the effectiveness and practical effect of the method of the present invention, the method of the present invention and several common networks are applied to the chrysanthemum quality data set in this embodiment. The specific comparison results are shown in Table 1, and the optimal performance values ​​corresponding to different evaluation indicators are highlighted in bold:

[0070] Table 1 - Performance comparison results of different networks on the chrysanthemum quality dataset

[0071]

[0072] It can be seen from the above experimental data that the method of the present invention can accurately evaluate the quality of chrysanthemums and has a high consistency with the subjective evaluation results of the human eye.

[0073] In order to verify the effectiveness of the algorithm provided by the present invention in practical applications, Figure 2 The chrysanthemum images of different quality levels are taken as examples, where (a), (b), (c), (d), and (e) are arranged from high to low quality. It can be seen from the figure that the prediction results of the overall network model for chrysanthemums of different quality levels are very close to the actual results;

[0074] The method of the present invention for removing multi-level feature fusion and the complete method of the present invention are respectively executed to obtain Figure 3 The correlation scatter plots of the predicted values ​​and the true values ​​are shown, where (a) is the correlation scatter plot without the multi-level feature fusion in the present invention; (b) is the correlation scatter plot obtained by the complete algorithm of the present invention. It can be seen that the SRCC and PLCC performance of the distorted image data set processed by the complete algorithm is much better. It can be intuitively observed that the quality prediction results of the complete algorithm are more consistent with the subjective evaluation results, there are fewer outlier samples, and the outlier range of the outlier samples is also smaller, which shows that the multi-level feature fusion in the present invention helps to provide the accuracy and stability of the model, and the prediction score has a better correlation with the subjective evaluation results.

[0075] In this embodiment, in order to verify the effect of meta-learning summarizing different prior knowledge, the present invention is trained on different data sets (or data set combinations), and its performance effect on the chrysanthemum quality data set is tested. The relevant performance index results are recorded in Table 2, where the types of data set distortion and the amount of data during training are arranged from left to right in descending order. It can be found from the table that the present invention has achieved good results in a variety of different cross-dataset tasks, and the fine-tuning effect will be better as the types of data set distortion are enriched.

[0076] Table 2 - Performance comparison results on the chrysanthemum quality dataset after meta-learning on different datasets

[0077]

[0078] In this embodiment, in order to verify the effect of meta-learning in reducing data dependence, the present invention is trained on different numbers of distorted chrysanthemum images, and its performance effect on the chrysanthemum quality data set is tested, and the relevant performance index results are recorded in Table 3. It can be found from the table that the present invention can still achieve good results while reducing the amount of training data, and meta-learning can effectively reduce the model's demand for data costs.

[0079] Table 3 - Performance comparison results of different proportions of training sets

[0080]

Claims

1. A chrysanthemum image intelligent screening method based on deep multi-scale fusion under a meta-learning framework, characterized in that: The steps include: S1: Through the constructed densely connected deep feature extraction module, three different levels of features of the distorted image are extracted, and the three different levels of features are spliced ​​as deep fusion features under different receptive fields; S2: Use the deep fusion features as the feature input of the constructed semantic analysis module based on Vision Transformer, analyze the mutual relationship and semantic information between multi-level deep features, and obtain semantic features; S3: The semantic features are passed through a linear connection layer to obtain the quality score; S4: The deep feature extraction module, the semantic analysis module and the linear connection layer are combined to form a backbone network. During the execution of steps S1 to S3, the constructed meta-learning framework is used to learn and simulate the parameter update mode of the backbone network. The distorted image dataset is used as a meta-training object, and the experience of evaluating the image quality rules is summarized from the distorted images. The meta-learning framework periodically learns each meta-task according to the learning step size and saves the learned prior knowledge. S5: Input the distorted image of chrysanthemum to be identified into the backbone network, and fine-tune it after integrating prior knowledge into the backbone network to obtain a model suitable for chrysanthemum quality evaluation, obtain the chrysanthemum quality evaluation result, and perform quality screening on the chrysanthemum image.

2. According to the method for intelligent screening of chrysanthemum images based on deep multi-scale fusion under a meta-learning framework according to claim 1, it is characterized in that: The deep feature extraction module in step S1 is composed of head feature extraction, 3 dense modules and 3 transition layers.

3. According to claim 2, the method for intelligent screening of chrysanthemum images based on deep multi-scale fusion under a meta-learning framework is characterized in that: The head feature extraction is composed of a convolution layer, a normalization layer, an activation function and a maximum pooling layer, which are used to preliminarily adjust the feature size.

4. According to claim 2, the method for intelligent screening of chrysanthemum images based on deep multi-scale fusion under a meta-learning framework is characterized in that: Three dense modules and three transition layers alternately complete the feature extraction task. The three dense modules are composed of 6, 12, and 24 dense layers respectively. Each dense layer is composed of two groups of convolutional layers, normalization layers, and activation functions to meet the extraction of current receptive field features. The input and output features of each dense layer are concatenated with channel features as the input of the next dense layer to reduce gradient vanishing and increase the richness of feature transfer. The transition layer after each dense module is composed of a convolutional layer, a normalization layer, an activation function, and average pooling to reduce the overall feature size and the number of channels.

5. According to claim 4, the intelligent screening method of chrysanthemum images based on deep multi-scale fusion under a meta-learning framework is characterized in that: The three transition layers at different levels will obtain features of 56*56, 28*28, and 14*14 sizes respectively. The 56*56 and 28*28 features are convolved to obtain 14*14 features. The 14*14 features at three different levels are channel-joined to obtain preliminary fused deep features, namely, deep fusion features.

6. According to claim 2, the method for intelligent screening of chrysanthemum images based on deep multi-scale fusion under a meta-learning framework is characterized in that: In the semantic analysis module of step S2, the PatchEmbedding part of the Vision Transformer is removed, and the semantic analysis module uses Position Embedding to annotate the position information of different channel features for subsequent analysis of the weight relationship between different features; the two-dimensional features are reduced to one dimension to form a serialized input, and the Encoder is used to analyze the feature semantics. The Multi-headAttention in the Encoder is used to assign Q, K, and V features to each sequence feature, and the weight mapping relationship between different features is obtained by weighting. On the basis of the attention mechanism, each position in the input sequence is allowed to pay attention to other positions in the sequence, and the weighted average is dynamically calculated as the output, so as to capture the dependency relationship in the sequence, thereby further obtaining the overall and local contextual relationship of different channel features, and finally obtaining the semantic information and weight relationship between different features.

7. According to claim 6, the method for intelligent screening of chrysanthemum images based on deep multi-scale fusion under a meta-learning framework is characterized in that: The process of obtaining weight relationship features by Multi-headAttention is as follows: softmax is equivalent to normalizing the obtained weights, d k Represents the dimension size:

8. According to claim 1, the method for intelligent screening of chrysanthemum images based on deep multi-scale fusion under a meta-learning framework is characterized in that: In step S3, the linear connection layer includes multiple channel connection layers to meet the mapping requirements of the overall channel from 1000 to 512, 512 to 256, 256 to 32, and 32 to 1, and normalization operations and activation functions are interspersed between channel mappings; the mean square error loss is selected as the loss calculation, and the Adam optimizer is designed to calculate the overall gradient change and perform back propagation to gradually obtain a model with better performance until convergence; wherein, the loss function is calculated as follows, y pre is the quality score predicted by the model, and y is the actual quality score corresponding to the image: Loss=||y pre -y|| 2 。 9. According to claim 1, the method for intelligent screening of chrysanthemum images based on deep multi-scale fusion under a meta-learning framework is characterized in that: In step S4, the meta-learning framework updates parameters by periodically learning the network. The specific process is as follows: the images with the same distortion type in the distorted images are classified into the same meta-task T i , the images in the meta-task are divided into the support set in a 4:1 ratio and queryset When the model is trained on both the support set and the query set, meta-learning will learn and summarize the prior knowledge; the same meta-task T i The support set in and queryset Divide into multiple subtasks and The rule of meta-learning summarizing prior knowledge is as follows: each time a single meta-task is trained, the framework analyzes some network parameters according to the set learning step size. When a set number of tasks or the current training stage are completed, meta-learning updates the accumulated prior knowledge to the parameter dictionary. When all training tasks are completed, the meta-learning framework completes the learning of the network model and summarizes the regional connections and semantic information of all types of images, forming prior knowledge with generalization and stability. Different learning rates ∈ 2 are applied to the two module parts and the linear connection layer during the meta-training process. fnet ,∈ Linear For network update, when the network trains the jth subtask, the parameter update process of the multi-channel feature extraction network and the inter-layer fusion network is: in, and They represent the parameter update results of the j-th support set subtask and the j-1-th support set subtask in the i-th meta-task respectively; the role of ε is to avoid the divisor being 0; and They represent the first-order moment estimate and the second-order moment estimate after bias correction in the j-th task, respectively. and The specific calculation process is as follows: The parameter update process of the linear regression part is the same as above, just change ∈ fnet The corresponding learning rate is replaced by ∈ Linear The corresponding parameters can be obtained; the learning method of meta-learning is controlled within the step size σ of the cycle, and the memory capacity of meta-learning will also change with the sequence j of each subtask; when training each subtask, from each meta-task T i The prior knowledge θ learned in i The change is defined as: When all meta-tasks T i After the training and learning of are completed, the prior knowledge θ containing all the distortion information is formed. prior .

10. The method for intelligent screening of chrysanthemum images based on deep multi-scale fusion under a meta-learning framework according to claim 9, characterized in that: In step S5, the prior knowledge θ obtained by meta-learning is prior Import it into the backbone network, and use the backbone network retraining process to quickly fine-tune the model used for chrysanthemum quality evaluation. In view of the efficiency and accuracy of the fine-tuning process, ∈ fnet and ∈ Linear Set to the same value ∈, that is, the module part of the backbone network and the linear connection layer use the same learning rate; Fine-tuning process, the process of the model obtaining the chrysanthemum quality score is: y pre =f(θ prior ,∈).

Images