IRVD-NAS: A lightweight backbone for infrared vehicle detection

This research introduces a neural architecture search (NAS) based backbone network called IRVD-NAS, which is designed for infrared vehicle identification (IRVD) in aerial photography. IRVD is a challenging task due to small target sizes, low image resolution, and complex backgrounds. To solve these problems, IRVD-NAS first searches for a lightweight yet effective backbone structure and further improves the results using a delayed down-sampling strategy. Unlike previous NAS methods, IRVD-NAS uses detection accuracy as the supervision signal, instead of classification accuracy, which is more favorable for the infrared vehicle detection task. Besides detection loss, the target-background pixel ratio is also used as a supervision signal to mitigate the interference caused by small vehicle targets and complex backgrounds. We suggest a down-sampling delay approach to enhance the detection of minuscule objects, based on the network architecture that was investigated. By giving more importance to detailed features at lower levels, this approach can greatly enhance the ability to detect small targets without requiring much computational resources. Extensive experiments on DroneVehicle dataset show that IRVD-NAS significantly improves accuracy and achieves real-time speed in the aerial photography infrared vehicle detection task.