README.md

Crack_detection_metal_part

A python flask app running on local host. Due to computation limitation, images if trimmed to 512x512 pixels.

Pre-requisite nstall requirements

pip install -r requirements.txt

Run the program

Some preprocessing to image

Structure of folders :

• Download the dataset from DataSet and extract the contents to the folder "YE358311_Fender_apron"
• Dataset Directory containing "normal" and "defect" folders
  • dataset ---> "/YE358311_defects/YE358311_Crack_and_Wrinkle_defect/"
  • dataset ---> "/YE358311_Healthy/"
• Destination Directory containing "train" and "test" folders
  • data ----> train ----> {"normal", "defect"} subfolders
  • data ----> test ----> {"normal", "defect"} subfolders

Training a simple CNN classifier (3 Conv + 1 FC)

• python3 main_train.py --- Start the training application server(local)
• http://127.0.0.1:5000/crack_detection_train --- Train the model
• python3 main_predict.py --- Start the testing application server(local)
• http://127.0.0.1:5000/crack_detection_test --- Opens an hmtl to upload the image and predict if Defective or Healthy

Accuracy Metrics

Validation accuracy and loss

• Training loss : 0.1695
• Training accuracy : 97%
• Validation/Test loss : 0.2475
• Validation/Test accuracy : 92.4%

Need for Improvements

• Next commit is to submit an andriod one page app to interact with the python flask API, showing upload, train and test functionality just like the webapp above.
• Image Preprocssing/Data Preparation
  • Since cracks are of less area as compared to image and noise, will introduce dropouts to improve accuracy
  • Preparing a object extraction module (For removing the rest of noise to improve accuracy)
  • Preparing a background color update module (For easy extraction of metal part after grayscale conversion)
• Trainng Improvements
  • Using a pretrained model such as VGG16 trained on ImageNet
  • Using model ensembles such as CNN+SVM (rbf kernel), Gauss filter+LBP+SVM(rbf kernel) etc that have proved improving accuracy