First Order Features Extraction from Thermal Images for Human State Recognition

Due to the common limitation of the human visual system, internal features of thermal images cannot be fully discovered. To overcome these drawbacks, a lot of studies analyzed the facial expressions corroborating the features extracted from thermal images with machine learning tools. This paper proposed the first-order features extracted from the image histogram and a binary classification performed by the k-nearest neighbor algorithm (K-NN). The proposed method tests separately the performance of different input data, firstly, the features were extracted from raw images and the robustness of the K-NN algorithm led to an accuracy of 92.6%. Secondly, the features were extracted sequentially from corrupted images with noise. The novelty of the proposed method consists of verifying the quality classification when the images have bad quality, in this case, the accuracy decreased to 70.7%, as a result, the classification of facial expressions is sensitive to the quality of images.