J. Eur. Opt. Society-Rapid Publ. 22, 43( 2026) 423
Figure
4. Network structure diagram of the improved YOLO-KMM model.
Figure
5. C2KD feature enhancement module fused with token statistics self-attention for ultra-small target weak feature enhancement and background noise suppression.
Figure 6. MU bottleneck block structure integrated with depthwise separable convolution for nonlinear feature capture and computational complexity reduction of infrared small targets.
computation. This improvement retains the original CSPNet split-fusion structure of the C3K2 module, ensuring compatibility with the overall YOLO11 architecture, and is particularly suitable for scenarios such as infrared small target detection that require capturing the features of weak and small-sized targets.
The Conv-M module described above is integrated as the core building block of the MU bottleneck within the
C3K2-MU structure. Specifically, Conv-M handles directional feature capture through asymmetric padding operations, while the MU mechanism applies nonlinear feature fusion. This synergistic design ensures that C3K2-MU simultaneously captures both directional characteristics( via Conv-M) and nonlinear feature correlations( via MU), providing comprehensive feature representation for ultrasmall infrared targets.