metallic

This is repo for our paper: A channel attention based deep neural network for automatic metallic corrosion detection.

project structure specification

preprocess.ipynb --- for image preprocessing

./dataset --- our dataset path

​ /train --- path of training dataset (ImageFolder in Pytorch)

​ /test --- path of test dataset

./utility --- helper module

metrics.py introduction

about the computation of metrics (defined by many evaluation metrics)

computeMetrics(Ypred:list, Ytest:list) -> {'acc':acc...}

output.py usage for raw image preprocessing

Overall, each raw image in the raw dataset contains a single metal sheet with the same level of corrosion or entirely non-corrosion. We label the metal area (foreground) by polygon in LabelMe. Then, we crop per image into patches by non-overlapped sliding window technique into C+1 folders (C corrosion levels + background class). Finally, we directly load the dataset with torchvision.data.dataset.ImageFolder. Notice that the evaluation is based on the image level rather than patch level (by re-mapping patches to their original position on the raw images).

In this step, the raw image will be fragmented in a sliding window manner, and the label of each patch are computed by the overlap with annotated bbox (annotated by LabelMe), i.e., foreground and different levels of corrosion.

three arguments for DataProcessor initialization

• file path of images and annotations
• folder of background images
• folder of foreground images

four arguments for process method for image fragmentation (sliding window)

• sliding window length
• sliding window width
• sliding window x-axis step size
• sliding window y-axis step size