Detecting Smoke in an Image Using Cascade Classifiers

摘要

This paper proposes smoke detection using image processing technique. It has been a challenging problem for a decade or two because of its variation in color, texture and shape. In this project a machine learning based approach is considered to solve this problem. The smoke detection problem is posed as classification problem. The solution is modeled as binary classification problem. Therefore, support vector machine (SVM) is considered for classification. In order to train and test the SVM classifier, positive and negative samples are collected. Two SVM classifiers are used in cascade. The first classifier detects the presence of smoke, if smoke presents in a given input image, the second classifier is used to locate the region of smoke in a given input image. The size of the window is fixed to 32x32 and slided across the entire image to detect the smoke in a region of the window. The model is trained a training dataset and using linear kernel as a kernel function. Subsequently, the model is tested with a Test dataset. The first SVM classifier achieved a training accuracy of 100% and testing accuracy of 92.5%. The second SVM classifier achieved a training accuracy of 96.5% and testing accuracy of 91.5%.