Optimal selection of fractal features for man-made object detection from infrared images

In this paper, a review of man-made object detection algorithms is presented based on various fractal features which are derived from the blanket covering method. These fractal features include fractal dimension (D), fractal model fitting error (FE), D-dimension area (K), multi-scale fractal feature related with D (MFFD), and multi-scale fractal feature related with K (MFFK). To choose the optimal fractal feature for man-made object detection from infrared images, a performance evaluation method for these algorithms is proposed in criterion of overlapped regions between ground truth and segmented image. The analysis and comparison of these algorithms are performed in terms of detection accuracy and computation cost. The results have revealed that different fractal features have different capability in discriminating between natural and man-made objects, and MFFK has the highest detection accuracy among all evaluated fractal features.

I. Introduction

Automatic detection of man-made objects in natural scenery from infrared images is an important and potential area in pattern based image analysis. The main difficulties are: 1) there is sometimes no significant difference between man-made object and its surroundings in thermal contrast; 2) infrared images usually have a high noise level and a low signal to noise ratio ; and 3) for small man-made objects, there is no geometric, spatial distribution, and statistical information to make use of [1]. Since fractal geometry was introduced by Mandelbrot to describe highly complex objects and natural scenery [2], it has been applied to detect man-made objects from complex natural scenery [3]–[6].