Artificial intelligence-based smart forestry aerial image sample data amplification method, device, platform and storage medium

Abstract

A method for augmenting sample data of smart forestry aerial images based on artificial intelligence (AI) is disclosed, comprising the steps of combining a WGAN-GP model with an ACGAN model to obtain a generative adversarial network model (WACGAN-GP); inputting original forestry aerial image sample data into the WACGAN-GP model to generate a first sample dataset; and simultaneously inputting the original forestry aerial image sample data into a VGG19 model to obtain a feature-extracted improved VGG19 model.

Description

Technical Field

[0001] The present invention belongs to the technical field of forestry inspection, and in particular relates to a method, device, platform and storage medium for amplifying sample data of smart forestry aerial images based on artificial intelligence. Background Art

[0002] When using algorithmic models to detect targets in forestry images, if the sample size is small, resulting in poor detection results, data augmentation is necessary. To obtain more data, the existing data can be adjusted accordingly, such as by flipping, rotating, and translating. The neural network model then treats these adjusted images as different images. Data augmentation is an important spatial solution to the problem of limited training data. It aims to expand the scale of training data, alleviate the overfitting problem of deep neural network models, and improve the performance and generalization ability of the model. Data augmentation has evolved into many methods, but because they always perform transformations on the original image, the improvement in classification performance is still very limited. Even the subsequent emergence of automated search data augmentation strategies only simplifies the strategy selection process and does not substantially improve the classification performance of the dataset. Summary of the Invention

[0003] One embodiment of the present invention is a forestry sample data augmentation method based on a generative adversarial network: combining a WGAN-GP model with an ACGAN model to obtain a generative adversarial network model WACGAN-GP; inputting original forestry aerial image sample data into the generative adversarial network model WACGAN-GP to generate a first sample data set;

[0004] The original forestry aerial image sample data is input into the VGG19 model to obtain a feature extraction improved VGG19 model;

[0005] Inputting the first sample data set into the feature extraction improved VGG19 model, and performing sample extraction on the first sample data set to obtain a second sample data set;

[0006] The data selection method ENN-IHT is adopted for the second sample data set. The data is selected using ENN-IHT, and the noise data is filtered out to obtain an enhanced sample data set. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which:

[0008] Figure 1Flowchart of a sample data augmentation method according to one embodiment of the present invention. DETAILED DESCRIPTION

[0009] Data augmentation can increase the number of samples in the training set, effectively alleviating model overfitting and making the training data as close as possible to the test data, thereby improving prediction accuracy. Sample augmentation can also force the network to learn more robust features, thus enhancing the model's generalization capabilities. For example, by applying a certain degree of occlusion to an image.

[0010] Generative Adversarial Network (GAN) is a new deep learning framework that can learn identification features from images and generate real samples. The images generated by GAN mainly depend on the model structure and the quality of the sample set. With the rise of generative adversarial networks (GAN), GAN can combine the high-dimensional features of existing data sets to generate images that are completely different from the original data sets. Compared with traditional data augmentation methods, it can provide more image features for neural networks. A carefully designed generative adversarial network can generate clear images that are similar to the original images. However, GAN also has problems such as unstable training, low quality of generated images, inability to generate images with specified labels, and noise data contained in generated images. Therefore, the present invention proposes a data augmentation method based on generative adversarial networks to address these problems, namely,

[0011] The WGAN-GP model can quickly and stably generate high-quality images, but it cannot generate images with specified labels. The ACGAN model can generate images with specified labels, but it cannot stably generate high-quality images. This paper combines the advantages of WGAN-GP and ACCGN to design a generative adversarial network model, WACGAN-GP, which incorporates gradient penalties and a new loss function. It also introduces a series of optimization strategies, including label smoothing, mini-batch discrimination, spectral normalization, and residual blocks, enabling WACGAN-GP to generate clearer and more distinguishable images and achieve better results.

[0012] According to one or more embodiments, a generative adversarial network model WACGAN-GP model framework is used to solve the problems of existing generative adversarial networks easily crashing during training, low quality of generated images, and inability to specify the generated image category. The forestry image sample data augmentation method includes:

[0013] 1) Combining the advantages of WGAN-GP and ACGAN, a generative adversarial network model WACGAN-GP is proposed.

[0014] 2) The WACGAN-GP model can generate high-quality images with specified labels to meet the needs of image data augmentation. Figure 1 shown.

[0015] Furthermore, in order to deal with the low-quality data that may appear in the samples generated by WACGAN-GP, a data selection method is proposed to clean up the noise samples to prevent them from affecting the classification and generalization capabilities of the model, including:

[0016] 1) The original dataset is trained on the improved VGG19 model to obtain a feature extraction model. VGGNet, a classic deep convolutional neural network, is a commonly used feature extraction model. Its deep network structure can extract deep features of the data. VGGNet is a deep network, and using the VGG19 model not only prevents overfitting during training, but also offers a deeper network than VGG16, resulting in stronger feature extraction capabilities. The improved VGG19 model removes the fully connected layers from the original model and adds improved fully connected layers.

[0017] 2) Extract features of the dataset generated by WACGAN-GP.

[0018] 3) Data selection uses a new data selection method called ENN-INT, which combines the advantages of both the ENN and Instance Hardness Threshold (IHT) algorithms. When using ENN-IHT for data selection, most of the noise in the original dataset is removed after the ENN selection. IHT effectively compensates for ENN's difficulty distinguishing boundary samples, thereby improving the quality of the generated dataset.

[0019] The embodiment of the present invention obtains an enhanced sample data set by collecting forestry image sample data, annotating all obtained sample data, and calculating weights and parameters of the WACGAN-GP algorithm.

[0020] The beneficial effects of the present invention include that an image data augmentation method based on generative adversarial networks and data selection is proposed, which can generate high-quality image data for different data sets, can be effectively applied to the field of image recognition, effectively expand the image data set, and improve the model classification performance. In the future, the network structure will continue to be optimized to enable it to run more efficiently, further adapt to unbalanced data sets, and optimize the data selection algorithm to enable it to perform adaptive parameter selection. By adopting WACGAN-GP, the present invention can make the training of GAN more stable and generate more realistic images. ACGAN can generate images with specified labels based on image labels. Combining the advantages of the two to generate high-quality images with specified labels has very important research significance for image data augmentation.

[0021] It is worth noting that although the foregoing content has described the spirit and principles of the present invention with reference to several specific embodiments, it should be understood that the present invention is not limited to the disclosed specific embodiments, and the division into various aspects does not mean that the features of these aspects cannot be combined. Such division is merely for the convenience of expression. The present invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A method for amplifying sample data of smart forestry aerial images based on artificial intelligence, characterized in that: The following steps are involved: Combine the WGAN-GP model with the ACGAN model to obtain the generative adversarial network model WACGAN-GP; Inputting the original forestry aerial image sample data into the generative adversarial network model WACGAN-GP to generate a first sample data set; Inputting the original forestry aerial image sample data into the VGG19 model to obtain a feature extraction improved VGG19 model; Inputting the first sample data set into the feature extraction improved VGG19 model, and performing sample extraction on the first sample data set to obtain a second sample data set; The data selection method ENN-IHT is adopted for the second sample data set. The data is selected using ENN-IHT, and the noise data is filtered out to obtain an enhanced sample data set.

2. The sample data amplification method according to claim 1, characterized in that: The ENN-IHT method combines the nearest neighbor (ENN) and instance hardness threshold (IHT) algorithms to filter noisy data and improve the quality of generated samples.

3. The sample data amplification method according to claim 1, characterized in that: The feature extraction improved VGG19 model is achieved by removing the fully connected layer in the original VGG19 model and adding an improved fully connected layer to enhance the feature extraction capability of forestry aerial images.

4. The sample data augmentation method according to claim 1, wherein: The WACGAN-GP model generates high-quality images with specified labels by introducing a residual network and a gradient penalty mechanism.

5. An artificial intelligence-based smart forestry aerial image sample data augmentation device, the device comprising a memory and a processor coupled to the memory, the processor configured to execute instructions stored in the memory, the processor performing the following operations: Combine the WGAN-GP model with the ACGAN model to obtain the generative adversarial network model WACGAN-GP; Inputting the original forestry aerial image sample data into the generative adversarial network model WACGAN-GP to generate a first sample data set; Inputting the original forestry aerial image sample data into the VGG19 model to obtain a feature extraction improved VGG19 model; Inputting the first sample data set into the feature extraction improved VGG19 model, and performing sample extraction on the first sample data set to obtain a second sample data set; The data selection method ENN-IHT is adopted for the second sample data set. The data is selected using ENN-IHT, and the noise data is filtered out to obtain an enhanced sample data set.

6. The sample data amplification device according to claim 5, characterized in that: The ENN-IHT method combines the nearest neighbor (ENN) and instance hardness threshold (IHT) algorithms to filter noisy data and improve the quality of generated samples.

7. The sample data amplification device according to claim 5, characterized in that: The feature extraction improved VGG19 model is achieved by removing the fully connected layer in the original VGG19 model and adding an improved fully connected layer to enhance the feature extraction capability of forestry aerial images.

8. The sample data amplification device according to claim 5, characterized in that: The WACGAN-GP model generates high-quality images with specified labels by introducing a residual network and a gradient penalty mechanism.

9. An artificial intelligence-based smart forestry aerial image sample data augmentation platform, the platform comprising a server having a memory and a processor coupled to the memory, the processor configured to execute instructions stored in the memory, the processor performing the following operations: Combine the WGAN-GP model with the ACGAN model to obtain the generative adversarial network model WACGAN-GP; Inputting the original forestry aerial image sample data into the generative adversarial network model WACGAN-GP to generate a first sample data set; Inputting the original forestry aerial image sample data into the VGG19 model to obtain a feature extraction improved VGG19 model; Inputting the first sample data set into the feature extraction improved VGG19 model, and performing sample extraction on the first sample data set to obtain a second sample data set; The data selection method ENN-IHT is adopted for the second sample data set. The data is selected using ENN-IHT, and the noise data is filtered out to obtain an enhanced sample data set.

10. A storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Images