IR-DETR: An efficient detection transformer with multi-layer feature fusion for infrared small targets

Abstract

Automatic target recognition is critical in infrared imaging guidance. Detecting small targets is challenging due to poorly defined silhouettes, low signal-to-noise ratio (SNR) and complex backgrounds. Traditional methods and deep learning approaches often struggle with SNR, target size, and insufficient feature extraction. To address these issues, we propose the IR-DETR network family based on DETR. Specifically, we design a multilayer efficient backbone (MEB) for efficient feature extraction of infrared small targets using a simple and efficient network structure. We then fuse local and global spatial features through the local window attention-based intrascale feature interaction (LWAIFI). Our proposed TripleRepC3 (TRC3) expands the model’s receptive field, which improves the detection accuracy of infrared targets while also reducing the model’s overall size. Additionally, we introduce the Adaptive Max-Sigmoid activation function to address the shortcomings of previous activation functions in small target detection. Finally, by incorporating the Normalized Wasserstein Distance (NWD) loss function, we further improve the detection performance of IR-DETR for small infrared targets. Compared to SOTA models on ATR dataset, NUDT-SIRST dataset and IRSTD-1k dataset, the IR-DETR network family achieved the best performance. The code will be released at https://github.com/Eason215xB/IR-DETR.