Research On Pet-cat Face Detection Algorithm

Visual object recognition is a very active topic in computer vision and pattern recognition area. Over the years, person has long been centered around by most of the research works, such as face recognition, gender recognition, and so on. Besides person, vehicle, character, etc. are also common object of study. While animals like cats and dogs, which can be seen everywhere in daily life, rarely appear in the field. In resent years,a pet boom is in the ascendant, animals have been paid more and more attention to and began to step into the field of visual object recognition.In this paper,we focus on the cat face detection which would be applied in digital camera, proposed a coarse-to-fine detection algorithm based on machine learning method. Firstly, we adopt Haar-like feature based Adaboost learning algorithm to train a coarse cat face detector, which can rapidly detect cat faces in a variety of scales. Tested on the Microsoft cat image database(including 10,000 cat images), this coarse-level cat face detector achieved a detection rate of 85% , while with a high false alarm rate of about 27%. In order to reduce false alarm, a HOG feature based SVM classifier was cascaded, this fine-level detector would classify the normalized candidate cat face area detected by the coarse-level detector, then make the final decision. After the final classification, the decrease in false alarm was about 92%.To train the coarse-level detector, the look-up table(LUT) type weak classifier based Gentle Adaboost algorithm was adopted. LUT type weak classifiers are able to reduce the number of Haar features and describe the distribution of positive and negative samples better than threshold type weak classifier. Compared to Real Adaboost, Gentle Adaboost increases the numerical stability. The combination of Haar feature and Adaboost in the coarse-level detector is used to get a high proceeding speed, avoiding time-consuming defects of the HOG based SVM classification method; while the fine-level detector make full use of the HOG based SVM classifier's high accuracy to make up the deficiency of high false alarm by Haar based Adaboost classification.