Is He/she Wearing Sunglasses?
I. Introduction Computer vision and face recognition are both the areas closely related to machine learning technique. In this project, we used the grey-scale face images donated by Tom Mitchell [1] to classify if the person in picture is wearing a pair of sunglasses. We are trying to discover significant features in face recognition projects.
II. Data format A total 640 grey-scale images are in .pgm format. They are collected from 20 volunteers. 32 different pictures are taken from each volunteer. During the image collection process, volunteers are requested to turn their heads to four different positions: up, left, right and bottom. They are also asked to express four different facial expressions: neutral, happy, sad, and angry. For each composition of positions and facial expressions, two pictures of the volunteer are taken, one with a pair of sunglasses and the other without. All these features are packed in the title of the picture files. A sample picture is shown in Fig. 1.
Fig.1 A sample picture in the images.
Each picture is in three formats: a full-resolution (128*120), a half-resolution image (64*60), and a quarter-resolution image (32*30). For each pixel, it is represented as an 8-bit unsigned integer number for the grey-scale. Images are in 20 different folders named after the volunteer name. For the baseline test, we use pixels as features for classification. We use the quarter-resolution image with 960 pixels, since a dataset with higher resolution will cause heap overflow problem in weka. There are 960 features and 640 instances totally.
III. Baseline performance To run a Baseline test on the data set, we built a python program to traverse the data folders to get image names and path, read the image pixels and tagged them as with glass or without glass. Then the data set was outputted into a .csv file, which can be read by weka. The dataset is shown in Fig. 2.
Fig.2 Dataset of the project In the baseline test, we ran a linear regression algorithm on the data set in weka. We guessed linear regression algorithm tends to work well in this project. This is because intuitively, when a person in the image is wearing sunglasses, the image will have more dark pixels. The result is shown in Fig. 3. Without any advanced computer vision algorithm and