FACE DETECTION AND LOCALIZATION USING DATASET OF TINY IMAGES Swathi Polamraju and Sricharan Ramagiri Department of Electrical and Computer Engineering Clemson University ABSTRACT: Being motivated by the fact that there are billions of images available to represent the world visually, we took up this project starting off with a small subset of images. This project deals with the process of face detection and localization making use of the dataset with tiny images. There are many algorithms available for face detection, of which most of the algorithms focus on developing a parametric method of recognition. This project, in contrast to the above algorithms focuses on developing a non-parametric method of face recognition. The data itself helps in solving the problem of recognition. The technical methods used in this paper are very simple and at the same time perform an effective search for the process of detection and localizing the face. INTRODUCTION: All the methods in object recognition have two components: the model and the data. Vast majority of the current research has been into the modeling part where in parametric methods for recognition are being developed. The research by Torralba etc. moves in the opposite direction, exploring how data itself can help to solve the problem. The main aim of the project is to show that face detection and localization can be performed with the non-parametric measures in a much simple way, provided a large dataset of images is available. When many images are available, simple indexing techniques can be used to retrieve images with object arrangements to the query image. For this purpose, they have created a database of 79 million 32x32 color images. Each of these images has been labeled with one of the several non-abstract nouns in English. That means all the images come under some category or the other like face, person, tree etc., it becomes very easy to find images visually close to the query image, containing images with similar objects in similar spatial configurations. There are many applications for this large dataset: Object detection, object localization, image colorization of gray scale images and detecting image orientations:METHODOLOGY: The main aim of this project is to implement face detection and localization using this database of tiny images. Face detection: The main goal here is to detect whether an image has a face or not. Since the database has images which contain people, we use it to detect the face. Even though the size and orientation may vary considerably, this algorithm can successfully detect the faces. For this purpose we collect votes from all the nearest neighbors. Since these are already labeled, we know the category of images which get maximum number of votes. Thus if these belong to face, we know that our query image also contains a face. We use a small subset of images from the 79 million dataset. When the person size is large, the performance is significantly better than when it is small as the picture becomes more constrained and detection becomes easier. Face localization: Once the face has been detected in the image, we now have to locate its position. For this the segmented image is used and the best scoring segments gives the location. Basic steps that were followed are: 1. Input the query image. 2. Segment the image using Split and Merge Segmentation. 3. Search for a best match for each of the segments in the query image with tiny images in database. 4. A match corresponds to the segment with least dissimilarity with the given image in dataset. 5. Increment the score for a particular dataset every time a match is found. 6. Detect the face by using dataset with highest score (one with the maximum number of matches). 7. Locate the position of face in input image based on the properties of selected segment. Detailed Algorithm is as follows: 1. Load the input image. 2. Apply Split and Merge segmentation to obtain all the segments present in the input image. 3. Compute region properties for all the obtained segments and thus obtain area for each segment. 4. Impose area constraint and eliminate all the insignificant segments from the segmented image. 5. Call CompareSegments function. 6. Load all the images present in the dataset. 7. Resample each of the segments to a fixed size (usually equal to the images in dataset 32x32). 8. Compare all the segments with each of the tiny images in thedataset by computing the Sum Of Absolute differences. 9. Select the segments with minimum dissimilarity i.e., which gives the minimum SAD. 10. Assign a score (to dataset) for each of the matches between the segments and the dataset. 11. Detect the face based on the highest score obtained corresponding to the tiny images present in the dataset. 12. Face localization: Once the face is detected, face is recognized using the region properties. 32x32 images in dataset EXPERIMENTAL RESULTS: Input Gray Image Segmented Image Extracted Segment Extracted Binary ImageFinal Image with face detected and localized MORE RESULTS CONCLUSION: We observe that the implemented method for face detection and localization is quite successful given a dataset of images. Actually we have taken a subset of the available database. If the number of images in the dataset increases, theprobability of getting better results increases. As we have observed from our experiments the threshold values chosen also affect the output considerably. Although initially 32x32 seemed to be very small size for detection, it gave decent results showing that, object and scene recognition can be reliably performed for images of size 32x32. Another advantage of this method is that simple dissimilarity measures can be used for effective detection and localization. Thus simple non-parametric methods combined with a dataset of images can be used for face detection and recognition purposes. Using better metrics may improve the