Small target detection in infrared image using convolutional neural networks

Infrared small target detection is an important research topic in the field of infrared image processing and has a major impact on applications in areas such as remote sensing, infrared imaging precise. Due to atmospheric scattering, refraction and the effect of the lens, the infrared detector to receive the target information very weak, it's difficult to detect the small target in complex background. In this paper, a novel small target detection method in a single infrared image is proposed based on deep convolutional neural network that is mainly using to extract the features of target, through the method can obtain more discriminative features of infrared image. Firstly, the off-line training of convolution kernel parameters using open data sets and simulated data sets, the result of preliminary training gives an initial convolution kernel, this step can reduce the time required for parameter training. Secondly, the input infrared image is preliminarily processed by the trained parameters to obtain the primary features of the infrared image, through the processing of the convolution kernel, a large number of feature information in different scales of the input image are obtained. Finally, selecting and merging the features, design the efficient characteristic information selection strategy, then fine-tune the convolution parameters with the result information, by merging the feature graph can realize the output of the result target image. The experimental results demonstrated that compared with existing classical methods, the proposed method could greatly improve the quality of the results, more importantly, our method can directly achieve the end-to-end mapping between the input images and target detection results.