Computer Vision using OpenCV

Posted by Pranjal And Srashti Vijay on May 29, 2018

computer-vision

Started working on Computer Vision using OpenCV. Project is primarily to detect and recognise faces in live video stream in realtime.

  1. Detected the faces in some short video clips, firstly using pre-trained deep learning face detection model and then using Haar Feature-based Cascade Classifiers. This was done to compare the performance and results of the 2 approaches. Based on this, I have concluded that we can perform fast and more accurate face detection with deep learning face detection model.
  2. Executed a program to recognize the faces in a static image. While performing face recognition, we have to be careful about our training data set, it should contain sufficient number of images of all the persons to be recognised.
    Used this code to recognize the faces detected by pre-trained deep learning face detection models and was successful in doing so.
  3. Extended the code for webcam face recognition and started a project to recognize the face of the employees. In this project, I have used pre-trained deep learning Caffe model to detect faces in the webcam and training images and recognize the faces in the webcam by using LBPH face recognizer. Firstly, I collected the images of the employees and detect the faces in the images to use them in the training data set. I have used 12 random images of each person in the training data set.
  4. Run the code several times to identify the employees . The accuracy was found to be very low. Referred multiple resources on the web to narrow down on the issue being faced and how to go about fixing the same
  5. After a lot of research and fixing, I identified that recognising static images is much more efficient with better results compared to images dynamically pulled from video stream. Also there were many missed/ incorrect recognition with the image recognition part when trying with multiple employees (up to 10) so reduced the number of employees to 2, where there was higher missed rate and worked to get to the root cause of the same.
  6. When I increase the number of training images for the 2 employees from 12 to about 40, I got a good accuracy. As the training images is the only way to train a face recogniser we have to be very careful about selecting the training images.

Srashti Vijay
(IIT Varansi)

Intern at CodeFire Technologies
Mentor
Pranjal Srivastava

Need experienced Indian professionals
whatsapp