Crack segmentation for concrete pipes

In this work, the configuration of cameras for novel inspection systems and the semantic segmentation of cracks in pipes are undertaken. The convolutional transformer architectures TransMUNetand DeepLabV3+ are used as the segmentation model. Additionally, a Dilated Residual Block and a Boundary Awareness Module are included to capture details of cracks and boundary features, respectively. The TransMUNet model has already been evaluated using two publicly available datasets: Crack500 and DeepCrack [2]. Furthermore, all DeepLabV3+ models have been trained with the Crack500 and DeepCrack datasets, in addition to TransMUNet. In addition, experiments were conducted to identify optimal settings for photographing the cracks in the pipes. In the experimentation process, a series of images were captured in a variety of settings, which were then utilized as a dataset to test the model’s performance on the concrete pipes. The optimal settings, determined by the experiments, include the distance from the camera to the wall of the pipe and the illumination angle. The initial testing phase revealed that the DeepLabV3+ model with a ResNet-50 backbone exhibited an IoU (Intersection-over-Union) of approximately 31% and 24% for the dry and wet conditions of the concrete surfaces, respectively. However, the performance was

Datasets

Weights

1. Crack500 dataset

The Crack500 dataset contains 500 images of size around 2000 × 1500 pixels taken by cell phones on main campus of Temple University. And each image was cropped into 16 non-overlapped image regions and only the region containing more than 1000 pixels of crack is kept. Through this way, the training data consists of 1,896 images, validation data contains 348 images, test data contains 1124 images. Download the Crack500 dataset from this link. Here images and masks are in the same folder, however they are distingushed by file endings (jpg and png, respectively).

2. DeepCrack dataset

The DeepCrack dataset is consist of 537 RGB color images with manually annotated segmentations. The images were divided into two main subsets: a training set with 300 images and a testing set with 237 ones. You can download the Deepcrack dataset from this link.

|-- DeepCrack
    |-- train
        |-- images
        |   |-- 7Q3A9060-1.jpg
            ......
        |-- masks
        |   |-- 7Q3A9060-1.png
            ......
    |-- test
        |-- images
        |   |-- 11125-1.jpg
            ......
        |-- masks
        |   |-- 11125-1.png
            ......

3. Custom dataset

Custom made dataset is availabe by request. This dataset is used to test the semantic segmentation model and obtain optimal illumination angle. Angles are 10, 20, 45, 75 and 90 degree. Working distances start from 12.5 cm until 60 cm.

Weights

You can upload them to the checkpoints folder and edit the file names, hyperparameters in config_crack.yml.

Training

python train_crack.py

It will run training for model and save the best weights for the validation set.

Testing

python evaluate_crack.py --output <path_to_dir>

It will represent performance measures and will saves related results in results folder. Plots will be added to the same directory as evaluate_crack.py. Test set of of the DeepCrack and Crack500 do not have angles so, in order to use test set of these dataset, following part should be commented out. data_path = config['path_to_testdata'] DIR_IMG = os.path.join(data_path) DIR_MASK = os.path.join(data_path) img_names = [path.name for path in Path(DIR_IMG).glob('.jpg')] mask_names = [path.name for path in Path(DIR_MASK).glob('.png')] img_names= natsorted(img_names) mask_names=natsorted(mask_names)

Acknowledgment

This paper is based on following repos:

• TransMUNet
  
• DeepLabV3+