Drive link :https://drive.google.com/drive/folders/1cNSw_eawYD-KAeNkLj7GVsGxxIL5mnke?usp=drive_link This drive link contains the zip file of my dataset
Approach : Now my dataset was UTK dataset of which I have attached the link .In this datset the females are labelled 1 and male 0. This dataset was of people's faces without any covers, so the major challenge was to train the model for masked faces. To tackle this problem I reasearched a bit about face detection models and came to know about haarcascade. Now firstly I created another dataset folder "Face_detected_dataset" in which I detected the faces and then cropped them out and saved in this folder. Then I crosschecked again using haarcascade and deleted the files in which there were no faces in the folder "Face_detected_dataset". Now I deduced that a face usually covers 40 percent of our face so I again created another dataset folder "Cropped_face_dataset" where I croped the bottom 40 percent of every image in our folder "Face_detected_dataset" .
Now finally I used "Cropped_face_dataset" as my dataset and trained my model according to this. Now I extracted the labels 1 or 0 from the "Cropped_face_dataset" and then added it to an np array and then I trained the model with these labels which are essentially the genders and the images. After this for testing I again used haarcascade and detected face then cropped the bottom 40 percent and then tried to predict the gender. After this I took approx 100 masked images and labelled them and then I tried to predict their gender and compared it to their labels and calculated the accuracy which is 65 percent .Then I impemented Gradio.I have added these test faces to my google drive.
HOW TO RUN CODE: First of download the UTK face dataset. Then downlaod the haarcascade file I have attached in the same deopsitory as your file in which you are running your code. Also please make sure all the libraries are isntalled.
UTKface dataset :https://susanqq.github.io/UTKFace/
Screenshots:
opencv.ipynb is my solution file
gender_classification_model.h5 is my model