SAYOLO: Spatial–Frequency Aware YOLO Network for Infrared Dim and Small Target Detection

Infrared dim and small target detection is crucial in fields, such as military reconnaissance and remote sensing. However, due to the extremely small target size, lack of texture, and low signal-to-noise ratio, the direct application of general-purpose object detection models leads to poor accuracy. To address these issues, we propose a spatial–frequency aware network for infrared dim and small targets, named SAYOLO. Different from simple adaptation of general-purpose architectures, this work starts from the characteristic of infrared dim and small targets manifesting as high-frequency signals in images, achieving effective migration of the YOLO network from the general object detection domain to the infrared dim and small target detection domain. Specifically, we design a position-sensitive enhancement module. Through the synergy of positional encoding and regional spatial masking, it addresses the inadequacy of existing attention mechanisms in modeling the positions of dim and small targets. To tackle the limitations of fixed-scale perception methods, we introduce a frequency-aware module. By combining a frequency-progressive dilation rate group strategy and dual-path attention, it achieves multiscale contextual perception of low-frequency background and high-frequency targets. Finally, to address the sharp drop in scale sensitivity, we design a dimension-stable intersection over union loss function, establishing a dynamic relationship between scale differences and loss penalties. Experiments show that SAYOLO delivers superior performance on multiple datasets. On the IST-A dataset where targets are extremely dim and small, the detection accuracy of SAYOLO improves by approximately 3.8% compared to the current best methods, highlighting the effectiveness and advancement of the proposed method for the specific task of infrared dim and small target detection.