Influence analysis of wide dynamic infrared image quantization on target detection performance

Abstract

Deep learning-based object detection networks outperform the traditional detection methods. However, they lack interpretability and solid theoretical guidance. To guide and support the application of object detection networks in infrared images, this work analyzes the influence of infrared image quantization on the performance of object detection networks. Firstly, the traditional infrared quantization methods and deep learning-based object detection networks are introduced, and the characteristics of these methods are analyzed. Then, the influence of four typical quantization methods on the performances of two object detection networks is compared, and the influence mechanism is analyzed through a cross-comparison experiment. The experimental results show that infrared image quantization is more helpful for learning the discriminative feature of the object/background for the object detection networks. Moreover, the feature difference of object/background caused by different quantization methods will seriously affect the performance of object detection networks. The research provides support for the application of deep learning-based detection networks in infrared scenes, which is of great significance.