Improving the Classifier Performance in Detecting People Based on Denoising Wavelet Transform

This paper presents the effects of the noise on the performance of the classifier in detecting people using Infrared (IR) camera, and then improve its performance by using denoising wavelet transformer techniques. The local binary pattern (LBP) detector is used in detecting person in IR images. The LBP features are extracted to train the classifier using a support vector machine (SVM). Experimentally, we find the classifier performs very poorly with a noisy image. Three wavelet functions (Harr, db2, and db4) are used in denoising process with different levels in an effort to find an efficient denoising method. Three type of noise models are added to the original test data set and three metrics namely- True positive Rate (TPR), False Negative per Frame (FNPF) and False Positive Per Frame (FPPF) are used to evaluate the performance of the classifier in detection process. The results show that denoising using db2 and db4 wavelet transforms with level 4 improve the classification results by removing successfully the three types of noise preserving the texture features of the original image which are used by the LBP detector.