An Enhanced Detection Method for Infrared Non-Cooperative Targets

To address the critical challenges of the low contrast and limited resolution of small targets and complex background interference in thermal infrared image-based non-cooperative target detection in airport environments， this paper proposes a universal enhancement detection algorithm specifically optimized for infrared non-cooperative targets. The proposed algorithm systematically improves the baseline model using four core plug-and-play modules. First， a small-scale detection head was introduced to significantly enhance the perception capabilities and localization precision of microscopic targets. Second， a space-to-depth convolution （SPDConv） module was employed to replace the traditional downsampling layers， which effectively preserved the fine-grained image features and mitigated the loss of critical details. Third， a reparameterized large-kernel inverted bottleneck （RepLKIB） module was designed. This module utilizes a parallel multi-branch structure to extract multi-scale features by incorporating a large-kernel convolutional branch to expand the receptive field. Furthermore， an inverted bottleneck structure was adopted to enhance feature representation， whereas structural reparameterization technology was applied to fuse the multi-branch training structure into a single-path inference structure， thereby balancing the accuracy and computational efficiency. Finally， an efficient intersection-over-union loss function was implemented to replace the complete intersection-over-union loss function. By decoupling the center and scale error terms， this approach significantly strengthened the regression gradients for low-quality samples， particularly small targets， thus improving the localization robustness and detection rates.Comparative experiments conducted on a self-constructed comprehensive infrared non-cooperative target dataset demonstrated the superior performance of the proposed algorithm. Building on the YOLOv10n （nano-scale） baseline， the mAP@［0.5：0.95］ metric increased substantially by 9.1 percentage points， reaching 94.0%. When compared with the larger YOLOv10m （medium-scale） model， the proposed model reduced the number of parameters and inference time to 70.3% and 75.7% of the original values， respectively， while incurring only a marginal 0.5% decrease in accuracy. A comparison with other versions of the YOLO series indicated that the algorithm achieved an accuracy comparable to that of medium-scale models while maintaining nano-scale inference speeds. Additional experiments using YOLOv11 as the baseline further validated the effectiveness and universality of the proposed modules across different architectures. In conclusion， this algorithm provides an efficient and reliable solution for infrared non-cooperative target detection in complex scenarios.