高级检索

基于深度神经网络的煤岩显微图像划痕检测与修复

Scratch detection and restoration of coal photomicrograph via deep neural network

  • 摘要: 煤岩显微组分的定量分析对客观评价煤的性质、质量进而实现煤炭的高效利用至关重要。然而,由于操作的不规范或刀具的破损,易造成煤光片中划痕的出现,影响进一步的自动化分析,而重新制作煤光片会造成人力和资源的浪费。鉴于此,为修复煤岩显微图像中的划痕区域,提出了基于深度学习的显微图像划痕检测与修复策略。首先,针对划痕检测设计了一种融合空间注意力与通道注意力的双注意力模型,充分挖掘煤岩显微图像的语义信息,并将其与U-Net语义分割网络融合,提高了划痕检测的准确性。然后,针对基于补丁匹配的图像修复算法会使煤岩划痕祛除区域与周围区域存在纹理差异的问题,设计了一种结合上下文注意力和生成对抗学习的煤岩显微图像修复网络。该网络自适应地为煤岩划痕区域填充合理内容,改善了图像修复质量。试验结果显示,划痕检测的平均像素准确度和平均交并比分别达到90.93%和83.95%,修复后的图像的峰值信噪比和结构性相似度分别达到43.29 dB和99.32%,相较于传统基于补丁匹配的算法,分别提高了8.76 dB和1.65%,验证了所提划痕检测与修复方法的有效性。

     

    Abstract: Quantitative analysis of coal maceral is crucial for objectively evaluating the properties and quality of coal, thereby enabling its efficient utilization. However, due to irregularities in operation or broken tools, it is easy to cause the scratch of the polished grain mounts, which affects the further automatic analysis, and the re-making of the polished grain mounts will cause the waste of manpower and resources. In view of this, in order to repair scratch areas in coal photomicrographs, a deep learning-based photomicrographs scratch detection and repair strategy is proposed. Firstly, a dual-attention model combining spatial attention and channel attention is designed for scratch detection, which fully mines the semantic information of coal photomicrographs and integrates it with U-Net semantic segmentation network to improve the accuracy of scratch detection. Then, to address the issue of texture differences between the scratch-removed areas and surrounding areas in image repair algorithms based on patch matching, a coal photomicrographs restoration network combining contextual attention and generative adversarial learning is designed. The network adaptively fills the scratched area of coal with reasonable content and improves the quality of image restoration. Experimental results show that the average pixel accuracy and mean Intersection over Union for scratch detection reached 90.93% and 83.95%, respectively, while the peak signal-to-noise ratio and structural similarity (SSIM) of the repaired images reached 43.29 dB and 99.32%, respectively. Compared to traditional patch-matching-based algorithms, these represent improvements of 8.76 dB and 1.65%, respectively, verifying the effectiveness of the proposed scratch detection and repair method.

     

/

返回文章
返回