煤矿智能化要求实现智能化开采，其中煤岩识别是实现无人化开采的核心技术之一，近年来基于图像的煤岩识别方法受到广泛关注。受采掘环境影响导致图像获取困难是制约煤岩图像识别的主要因素之一，针对煤岩图像数据短缺的问题提出解决方法并增加数据量，基于在单张图像上训练的生成式对抗网络提出Var-ConSinGAN模型并构建样本生成与特征迁移框架。ConSinGAN模型生成的图像虽然清晰但对于煤岩图像生成来说仍然缺少多样性，改进模型的训练方式和图像重构方法可提高生成煤岩图像的多样性。模型采用金字塔结构，使用多阶段训练方法，每个阶段训练不同尺度的图像，可以生成任意数量图像。在模型中提出了新的图像尺寸变换方法用来生成分辨率不同的重构图像，同时采用曲线函数使每一阶段的迭代次数逐步增加，之后对第1阶段用单张图像生成的结果使用基于辅助分类器的条件生成式对抗网络进行特征迁移。新图像重建方法保持重建图像高分辨率阶段的密度大于低分辨率阶段的密度，新训练迭代函数在减少模型学习图像结构的次数的同时增加模型学习图像纹理细节的次数。新的训练迭代函数可以减少模型迭代的总次数，从而减少模型训练时消耗的计算资源。模型需要输入一张符合正态分布的噪声图片，经过训练，生成器输出满足真实图像分布的生成样本。实验在500张煤岩图像上进行，每张图像生成400张仿真图像，并使用SSIM指标对生成图像的亮度、对比度、结构等3方面进行测评。其中结构性强的煤岩图像其SSIM指标值很低，反之，结构性弱的煤岩图像其SSIM指标值较高。根据实验结果可得结论:Var-ConSinGAN模型缓解了原始GAN在数据不足时无法训练和ConSinGAN生成的煤岩图像具有明显“边框”感且多样性不足的问题;Var-ConSinGAN与ConSinGAN相比模型性能提高了13.8%;特征迁移的实验结果显示不同种类的煤岩之间学习到了彼此没有的特征;增加了煤岩图像数据量。

The intelligentization of coal mine requires the realization of intelligent mining.Thereinto,automatic coal and rock recognition is one of the core technologies for unmanned mining.In recent years,image-based identification of coal and rock has been widely concerned.The difficulty of image acquisition due to the influence of mining environment is one of the main factors restricting coal and rock image recognition.To solve the problem of coal and rock image data shortage and increase the amount of data,this paper proposes the Var-ConSinGAN model based on the Generative Adversarial Networks(GANs) trained on single image and constructs the framework of sample generation and feature migration.Although the images generated by ConSinGAN model are clear,there is still a lack of diversity for coal and rock image generation.The improved training method and image reconstruction method of the model can improve the diversity of coal and rock image generation.The model in this paper adopts a pyramid structure and uses a multi-stage training method.Each stage trains images of different scales,and any number of images can be generated.A new image size transformation method is proposed in the model to generate reconstructed images with different resolution.At the same time,a curve function is used to gradually increase the number of iterations in each stage.Then the results generated from the single image in the first stage are migrated using Conditional Generative Adversarial Networks based on auxiliary classifier.The new image reconstruction method maintains that the density of the reconstructed image at the high resolution stage is higher than that at the low resolution stage.The new training iteration function reduces the number of times the model learns the image structure while increasing the number of times the model learns the image texture details.The new training iteration function can reduce the total number of model iterations,thus reducing the computational resources consumed during model training.The model needs to input a noise image that conforms to the normal distribution.After training,the generator outputs a generated sample that meets the real image distribution.The experiment was carried out on 500 images of coal and rock,each image generated 400 simulation images,and SSIM indicator was used to evaluate the brightness,contrast and structure of the generated images.Among them,the SSIM indicator value of coal and rock images with strong structure is very low,on the contrary,the SSIM index value of coal and rock images with weak structure is relatively high.According to the experimental results,it can be concluded that the Var-ConSinGAN model alleviated the problem that the original GAN could not be trained when the data was insufficient and the coal and rock images generated by ConSinGAN had obvious sense of“border”and insufficient diversity.Compared with ConSinGAN,Var-ConSinGAN improves the performance of the model by 13.8%.The experimental results of feature migration show that different types of coal and rock have learned features that do not exist in each other.The amount of coal and rock image data is increased.

1 生成式对抗网络

   1.1 模型框架

   1.2 基本原理

2 基于GAN的煤岩样本生成模型

   2.1 金字塔结构

   2.2 多尺度架构

   2.3 图像尺寸调整

   2.4 多阶段训练方式

   2.5 动态变化的迭代次数

   2.6 特征迁移

3 训练过程和损失函数

4 实验结果及分析

   4.1 无条件煤岩图像生成

   4.2 对比实验

   4.3 模型性能

   4.4 特征迁移

5 结论与展望

   5.1 结论

   5.2 展望