Client: Mr. Saurabh Parmar
Short Descp:
What I dealt on this project was to design quite good system to recognize faces under dark and dim light condition. Until now, light condition around objects (faces) to recognize is still an open problem. Many scientist is still observing better way to resolve it. This is a master thesis project. Accuracy and recognition speed are main issue to consider in the development. Improvement has to be developed based on the two base papers. So, the idea is not too choose or design new method for feature extraction and recognition, but more to design scheme for pre-processing steps. When we are dealing
with such problems, where lighting quality becomes big issue, we can simply normalize the image contrast into new contrast frame, so every time we input new image with different contrast the system would be able to recognize the face. Some papers used Gamma correction (monitor concept) to be implemented to whole image at one time. I think the best solution is to implement contrast enhancement to small window. In this research, window size based on Quadtree and Fix Sized has been tested. The best result is by using Fix Sized window but threshold of local contrast that i used is based on entropy value of each image. For feature extraction and classification, i use simple methods which are LBP and KNN. I this research, by using 45 images in training and 180 images for testing we can cover faces up to 97% of for all dark light and dim light categories.
with such problems, where lighting quality becomes big issue, we can simply normalize the image contrast into new contrast frame, so every time we input new image with different contrast the system would be able to recognize the face. Some papers used Gamma correction (monitor concept) to be implemented to whole image at one time. I think the best solution is to implement contrast enhancement to small window. In this research, window size based on Quadtree and Fix Sized has been tested. The best result is by using Fix Sized window but threshold of local contrast that i used is based on entropy value of each image. For feature extraction and classification, i use simple methods which are LBP and KNN. I this research, by using 45 images in training and 180 images for testing we can cover faces up to 97% of for all dark light and dim light categories.
Snapshots:
Fig 1. Face was successfully detected on highest dimness
Fig 2. Face was successfully detected on medium dimness
Fig 3. Face was successfully detected on lowest dimness
Fig 4. Time consuming for preprocessing, feature extraction and classification
No comments:
Post a Comment
Terimakasih telah membaca artikel dalam blog ini...