This project is using a dataset of images of different people either wearing masks or not. It is then used to create a deep learning model and check the accuracy and validation score.
SSD stands for Single Shot Multibox Detector. It is a technique that is used to detect objects in images using a single deep neural network. Basically its used for object detection in an image. By using a base architecture of VGG-16 Architecture, SSD is able to out perform other object detectors like YOLO and Faster R-CNN in both speed and accuracy. The architecture of SSD is given in the figure below. Training a SSD model from scratch will require a lot of data, so here I am using Caffe Face Detector Modelin OpenCV. This pre-trained model is part of the OpenCV library (version 3.3 onwards). It has been uploaded here for convenient usage.
The visualization tells us that the Number of Mask images > Number of Non-Mask images, so this is an imbalanced dataset. But since we are using a SSD pretrained model, which is trained to detect non-mask faces, this imbalance would not matter a lot. Along with that we are using keras preprocessing techniques to artifically increase our dataset size such as by flipping images horizontally, rotating them and etc.
I trained the model for 100 epochs and below I will discuss the results of accuracy and loss.
I achieved accuracy on training set: 96.86 % and on validation set: 96.87%. Please see the attached picture below.
I achieved loss on training set: 0.0791 and on validation set: 0.117. Please see the attached picture below.
The test dataset has 1698 images and to evaluate the model I have taken a handful of images from this dataset as there are no labels for faces in the dataset.
- Gamma Correction for making the image appear in more light.(Gamma = 2)
- blobFromImage creates 4-dimensional blob from image. Optionally resizes and crops image from center, subtract mean values, scales values by scalefactor, swap Blue and Red channels.
- The blob is passed through the SSD network and detections are made with some confidence score.
- Define a threshold confidence score, above which the detection will be considered as a candidate of being a face. (In this case confidence threshold = 0.2)
- All the detections that qualify the confidence score are then passed to the architecture for classification into mask or non-mask image.
The results are as follows:
By analyzing the results it can be observed that the whole system works well for faces that have spatial dominance i.e. in image at (1,1), (1,2) and partially for (2,1). It fails partially in case of (2,1) because it didnot detect the masked face of Mona Lisa. In (3,2) it fails to detect the third and fourth lady because their faces are small, blurrd, partially visible and occupy less space in the overall image. To get better results, different image preprocessing techniques can be used, or confidence threshold can be kept lower, or one can try different blob size.