Input: Cropped face or Non Cropped face images.

Detailed explanation: My project is face authentication. I am using OpenCV face detector to crop face from images. I am using OpenCV implementation of LBP for face recognition. My face recognition system works good when the the test images are not much variant (illumination) w.r.t training images. The constraint is, I don't have many variants during training; so, I need to categorize the images based on light/shadow and need to label each image . So that similarity score threshold value can be set dynamically based on the labels. I am not getting the desired results with the Illumination normalization algorithms (like Difference of Gaussian) and with the static threshold values for all variants of images . So, I need to know how to identify the quality(illumination) of images, to check the possibility of dynamic thresholds.

Any guidance on any of my above said process are appreciated and helpful to me.


1 Answer 1


It sounds like you better do it using Clustering approach.
The way I'd approach it is by creating a metric between images based on their L channel of the LAB Colorspace histogram.

There are many metrics for distributions.

Once you have that metric use a clustering algorithm which supports out of sample extension (K-Means, GMM, HDBSCAN).

This is a simple replacement to having a classifier. Yet a classifier might require more data and labeled data.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.