Skip to content

A comprehensive, two-tiered deep learning approach designed for precise object detection and segmentation in electron microscopy (EM) images.

License

Notifications You must be signed in to change notification settings

xiazeyu/DT_SegNet

Repository files navigation

Accurate identification and measurement of the precipitate area by two-stage deep neural networks in novel chromium-based alloys

paper doi open access paper license

code doi code license

About The Project

The performance of advanced materials for extreme environments is underpinned by their microstruc- ture, such as the size and distribution of nano- to micro-sized reinforcing phase(s). Chromium-based superalloys are a recently proposed alternative to conventional face-centred-cubic superalloys for high-temperature applications, e.g., Concentrated Solar Power. Their development requires the de- termination of precipitate volume fraction and size distribution using Electron Microscopy (EM), as these properties are crucial for the thermal stability and mechanical properties of chromium superal- loys. Traditional approaches to EM image processing utilise filtering with a fixed contrast threshold, leading to weak robustness to background noise and poor generalisability to different materials. It also requires an enormous amount of time for manual object measurements. Efficient and accurate object detection and segmentation are therefore highly desired to accelerate the development of novel materials like chromium-based superalloys. To address these bottlenecks, based on YOLOv5 and SegFormer structures, this study proposes an end-to-end, two-stage deep learning scheme, DT- SegNet, to perform object detection and segmentation for EM images. The proposed approach can thus benefit from the training efficiency of Convolutional Neural Networks at the detection stage (i.e., a small number of training images required) and the accuracy of the Vision Transformer at the segmentation stage. Extensive numerical experiments demonstrate that the proposed DT-SegNet significantly outperforms the state-of-the-art segmentation tools offered by Weka and ilastik regard- ing a large number of metrics, including accuracy, precision, recall and F1-score. This model will be a meaningful tool for accelerating alloy development and microstructure examination.

Getting Started

Quick Start

You can use the Inference_Colab.ipynb (Colab link) to perform online inference. Or, use following the instructions in Finetune_Colab.ipynb (Colab link) to perform online finetune.

Hardware requirement

  • Operation System: Windows or Linux
  • Platform: AutoDL / Google Colab Pro
  • GPU: NVIDIA RTX A5000
  • Google Drive space: 10GB

Software requirement

  • Programming language: Python (3.8 or higher)
  • Package Management: Anaconda (Miniconda recommended)
  • Machine Learning Framework: PyTorch and PaddlePaddle

Anaconda environment for labelling on Windows system is in dtsegnet.yaml. The environment can be restored by executing conda create --name dtsegnet --file dtsegnet.yaml in the console.

Two machine learning frameworks need to be installed following the tutorials on their websites. The necessary environment for training and inferring is stored as a pip requirement file in 1_Detection_Model/requirements.txt and 3_Segmentation_Model/requirements.txt.

All the requirements for training and inferring will be installed in the 0_Prepare.ipynb notebook.

Dataset

All data for this project are stored in the Dataset/ folder. All images are numbered for the DT-SegNet pipeline, and the data mapping is stored in dataset_mapping.csv.

The dataset contains the original image, segmentation label and detection label. Detection labels can be used directly for the detection network, but the segmentation label needs to be cropped using codes provided in the notebook before delivering to the segmentation network.

The detection dataset is separated into three sets: test, train, and val.

The segmentation annotation is stored in Dataset/segmentation_labels/. The Regions of Interest for the segmentation network with their annotations will be generated before the segmentation stage.

Implementation

0 Prepare

Follow the cells in 0_Prepare.ipynb to prepare required environments.

0 Label the dataset

0_Labelling_Tools/ contains tools of scripts to label the dataset. The user should follow the following process to label the dataset.

  • Execute conda create --name dtsegnet-label --file 0_Labelling_Tools/dtsegnet-label.yaml to import the Anaconda environment for labelling.
  • Execute conda activate dtsegnet-label to activate the labelling environment.
  • Download the model for EISeg labelling from https://paddleseg.bj.bcebos.com/eiseg/0.5/static_hrnet18s_ocr48_aluminium.zip.
  • Execute python 0_Labelling_Tools/0_EISeg/exe.py to start labelling for the segmentation stage.
  • Execute python 0_Labelling_Tools/3_gray2pseudo_color.py <Dataset/label> to convert the grey-scale segmentation labels to pseudo colour annotation images. <Dataset/label> should be replaced by the path of segmentation annotations generated by EISeg.
  • Execute python 0_Labelling_Tools/1_Segmentation_Label_Flood_Fill.py <Dataset/label> <Dataset/Detection_Label> to generate the detection labels from the segmentation labels. <Dataset/label> should be replaced by the path of segmentation annotations generated by EISeg, which contain files like: 1.png, 1_cutout.png and 1_pseudo.png. <Dataset/Detection_Label> should be replaced by the output folder for detection labels.
  • Execute python 0_Labelling_Tools/2_labelImg/labelImg.py to finetune the detection labels.

1 Train the model

Follow the cells in 1_Train.ipynb to train the detection model. The trained detection model will be stored in <Google Drive>/DT-SegNet/Detection_Model_Output. it will also automatically generate the dataset for the segmentation network. The generated segmentation dataset will be compressed and stored in <Google Drive>/DT-SegNet/Segmentation_Dataset.zip. The trained segmentation model will be stored in <Google Drive>/DT-SegNet/Segmentation_Model_Output.

2 Inference

Follow the cells in 2_Inference.ipynb to infer using DT-SegNet. The output from the detection model will be stored in <Google Drive>/DT-SegNet/Detection_Output. The output from the segmentation model will be held in <Google Drive>/DT-SegNet/Segmentation_Output. The original-size segmentation mask will be stored in <Google Drive>/DT-SegNet/Output.

3 Validation

Follow the cells in 3_Validation.ipynb to validate the trained models.

4 Analysis

Follow the cells in 4_Analysis.ipynb to analyse the trained models. To compare software and algorithms, we performed experiments on Weka trainable segmentation and Ilastik pixel classification. The notebook uses our output in Output/.

Main Models

Our best-trained models are stored on the Github release page.

Machine-friendly download links are also provided for automatic scripts:

https://github.com/xiazeyu/DT_SegNet/releases/latest/download/detection.pt
https://github.com/xiazeyu/DT_SegNet/releases/latest/download/segmentation.pdparams

Output

The Output/ folder holds this project's output images and NumPy metrics, including results from DT-SegNet, Weka and ilastik. Each experiment has an output in two different formats: the .png image output and the .npy NumPy matrix output. Each file in the folder is named by <test id>_<software used>_<algorithm>.

Software used Algorithm Remark
groundtruth groundtruth Manually annotated
ilastik LDA ilastik LDA (scikit-learn)
ilastik RF ilastik Random Forest (scikit-learn)
ilastik SVC ilastik SVM C-Support (scikit-learn)
weka FRF Weka hr/irb/fastRandomForest/FastRandomForest
weka MLP weka/classifier/functions/MultilayerPreceptron
with trainingTime=100 and validationSetSize=20
PaddleSeg unet U-Net
PaddleSeg unet_3plus UNet 3+
PaddleSeg deeplabv3p_resnet50 DeepLabV3+ with ResNet 50 Backbone
PaddleSeg B0 SegFormer B0
PaddleSeg B1 SegFormer B1
dtsegnet DT-SegNet DT-SegNet, with overlapping ROIs joined

Reference

@article{xia2023Accurate,
 author = {Zeyu Xia and Kan Ma and Sibo Cheng and Thomas Blackburn and Ziling Peng and Kewei Zhu and Weihang Zhang and Dunhui Xiao and Alexander J Knowles and Rossella Arcucci},
 copyright = {CC BY-NC 3.0},
 doi = {10.1039/d3cp00402c},
 issn = {1463-9076},
 journal = {Physical Chemistry Chemical Physics},
 keywords = {},
 language = {English},
 month = {6},
 number = {23},
 pages = {15970--15987},
 pmid = {37265373},
 publisher = {Royal Society of Chemistry (RSC)},
 title = {Accurate Identification and Measurement of the Precipitate Area by Two-Stage Deep Neural Networks in Novel Chromium-Based Alloy},
 url = {https://pubs.rsc.org/en/content/articlelanding/2023/CP/D3CP00402C},
 volume = {25},
 year = {2023}
}

License

MIT License. More information see LICENSE

Contact

Zeyu Xia - [email protected]

Kan Ma - [email protected]

Sibo Cheng - [email protected]

About

A comprehensive, two-tiered deep learning approach designed for precise object detection and segmentation in electron microscopy (EM) images.

Resources

License

Stars

Watchers

Forks

Languages