Mid Sweden University

Image classification is an image processing technique used to separate different categories in a dataset according to unique image information. It has many fields of application, such as medicine, security monitoring, component defect detection, people counting, license plate recognition and robot positioning. There are many challenges in the field of image classification. For example, poor image quality leads to poor results. Moreover, there is need to study how to improve the recognition rate, reduce the amount of calculation and strengthen the robustness [1].

Classification algorithms are tested for visual images in outdoor environments. However, vision sensors are sensitive to light. When there is no light, the signal to noise ratio in the visual images will be poor. This in turn could lead to false classification. To circumvent this problem, a sensor of another spectral band can be used, making sensing independent of lighting conditions. An example of this is thermal images which can be created by sensing the infrared radiation from objects. To evaluate the classification accuracy of different algorithms based on different image types, we developed five datasets: thermal raw images, visual raw images, thermal binary images, visual binary images and fused images. In this study, each dataset includes two different objects, a human and a bicycle. The classification algorithms used for investigations include HOG, SIFT and SURF. The first high level analysis is performed on a personal computer (PC). After analysis, SIFT and SURF algorithms were implemented on an embedded platform, NVIDIA TEGRA TK1.

The results show that thermal raw images is better for object classification compared to visual raw, thermal binary, visual binary and fused thermal-visual images. Regarding classification algorithms, the SIFT algorithm performs better with 93.3% accuracy compared to HOG and SURF which have 86.4% and 91.7% respectively.