Robust Small Target Detection Under Intense Noise Based on Extreme Learning Machine

The robust detection of dim small targets is the key problem in infrared search and track systems. However, the intense noise in infrared image has similar appearance and intensity distribution with the dim small target. This may lead to terrible probabilities of detection and serious false alarm rates for traditional small target detection methods. To cope with the disturbance of intense noise, a novel small target detection method based on the cascade sparse histogram feature and extreme learning machine is proposed in this paper. Firstly, top-hat transformation based pre-processing is applied on the infrared image to eliminate most of the background clutter. The top-hat based pre-processing can reduce the computation burden for the subsequent extreme learning machine based classification phase, however, some intense noise might still remain in the pre-processed image. Then, a novel descriptive feature namely the cascade sparse histogram is proposed to describe the characteristics of the target and the noise. Finally, the feature vectors of target and noise are discriminated by the extreme learning machine so the target could be detected. Extensive experiments show that the proposed method not only has outstanding small target detection performance under different backgrounds but also works more stably against intense noise compared with conventional baseline methods.