Oriented Infrared Vehicle Detection in Aerial Images via Mining Frequency and Semantic Information

Abstract

Infrared vehicle detection based on aerial images has significant applications in military and civilian fields for the perception ability under low-light and foggy conditions. However, it remains challenging due to the following characteristics. First, infrared textures and edges are blurred, implying a plenty of low-frequency signals and a shortage of detailed descriptions. Second, objects in infrared images present different patterns depending on their thermal radiations, which hamper the feature extraction of convolution kernels. Third, infrared images lack color information, which means fewer features for classification and regression can be used. Inspired by cognitive neuroscience that humans perceive the entirety from low-frequency information and discern details from high-frequency information, we devise a new framework for oriented infrared vehicle detection called I <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> MDet (Infrared Information Mining Detector) to tackle the above challenges. It consists of two significant designs: the kaleidoscope module and the semantic feature supplement module (SFSM). In the kaleidoscope module, we explore the effect of kernel sizes and dilation rates on frequency information mining with kaleidoscope-like equivalent kernels. Features in this module are extracted by adaptive involution operators instead of convolution kernels to deal with multiple patterns. The SFSM provides the network with features beneficial for classification and regression. On the one hand, the network is guided to learn more meaningful features under semantic supervision. On the other hand, features output by the SFSM supplement the detection head with semantic information. Experimental results on the public dataset DroneVehicle demonstrate that our proposed approach achieves outstanding performance on oriented infrared vehicle detection.