This project aims to determine if an image is blurred or not with a blurriness score along with brightness & contrast score. This project is specifically customized for automotive industry.
Techinical quality assessment : Low level degradations such as noise, blur, compression artifacts
Aesthetic quality assessment: semantic level characteristics such as emotions and beauty in images
You can play around here: https://share.streamlit.io/nikhilgunti/image-blur-detection/main/streamlit.py
Data: 20k clear images and 100k blurred images(generated synthetically using input clear images through Simple, Box and Gaussian blur functions)
Label: Label '0' as Not Blur and Label '1' as Blur
Scores: Blurriness, Brightness, Contrast
Features: Maximum and Variance values of Laplacian & Sobel filters
Methodology:
- Read the input image
- Convert the RGB image to grayscale mode
- Compute focus measure using different operators like Laplacian, Sobel, etc.
- If the computed value is less than the threshold, then the image is labeled as Blur
Model: Of all the trained classification models, I chose Random Forest classifer based on F1 score metric(imbalanced data)
Deployment: I have deployed the application in Streamlit Cloud. You can either select an image to test from the list or upload one of your own
Data prepartion : blur_preprocess.ipynb
Model inference : inference.py
Streamlit application : streamlit.py
img = cv2.imread(img_path)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
#Extracting Laplacian filter and Sobel filter values
lapvar, lapmax = varMaxLaplacian(gray)
sobvar, sobmax = varMaxSobel(gray)
#Label and score prediction based on pre-trained random forest model
class_label = ["Blur" if model.predict([[lapmax, lapvar, sobmax, sobvar]])==1 else "Clear"]
blur_score = model.predict_proba([[lapmax, lapvar, sobmax, sobvar]])
#Label and Score formatting
class_label = ''.join(class_label)
blur_score = str(round(blur_score[0][1],2))
