红外小目标检测因其在天基预警、海事救援等关键领域的战略价值而备受关注，但目标像素占比极小、信噪比低、背景复杂等特性使其成为极具挑战性的视觉任务。现有深度学习方法虽显著优于传统模型，但其普遍采用的交并比类损失函数对预测目标的绝对尺度及空间位置变化缺乏敏感性，导致模型难以实现像素级精确定位。针对上述问题，本文提出一种基于多尺度空间损失和形态学特征的红外小目标检测方法。首先，设计绝对空间损失函数，通过引入基于面积差异的自适应动态权重增强对目标尺度的感知能力，并基于极坐标系构建径向-角度惩罚项以精细化约束中心点定位误差；其次，在U-Net解码器中构建轻量级多尺度预测头结构，将AR损失同步施加于不同分辨率层级的预测输出，实现从粗到精的层级化监督；最后，构建训练-测试双阶段形态学增强策略，训练阶段采用基于池化的可微形态学算子嵌入结构先验，测试阶段通过开闭运算修正预测结果的连通性。在IRSTD1k数据集上，所提方法在交并比（IoU）、检测率（P_d）和虚警率（F_a）等指标上分别达到67.59%、93.02%和9.034×10^-6，较现有主流方法DNANet分别提升1.88%、1.18%，虚警率降低48.7%，在计算效率与检测精度间实现了更好平衡。

Infrared small target detection has attracted significant attention due to its strategic value in the key fields such as space-based early warning and maritime rescue.However，the extremely small pixel size，low signal-to-noise ratio，and complex background characteristics make it become a highly challenging visual task.Although the existing deep learning methods significantly outperform the traditional models，the intersection-over-union loss functions commonly used in the existing deep learning methods lack sensitivity to the absolute scale and spatial position variations of predicted targets，resulting in difficulties in achieving pixel-level precise localization and becoming a bottleneck for further performance improvement.To address these issues，this paper proposes an infrared small target detection method based on multi-scale spatial loss and morphological features.Firstly，an absolute-spatial ration （AR） loss function is designed，which enhances the perception of target scaletarget scale by introducing an adaptive dynamic weight based on area differences，and the radial-angular penalty terms in a polar coordinate system are constructed to refine the localization constraints of the center point.Secondly，a lightweight multi-scale prediction head structure is constructed in the U-Net decoder，thereby applying AR loss synchronously to the prediction outputs at different resolution levels to achieve a coarse-to-fine hierarchical supervision.Finally，a dual-stage morphological enhancement strategy for training and testing is constructed，embedding the structural priors via pooling-based differentiable morphological operators during training and correcting the connectivity of the predicted results through opening and closing operations during testing.On the IRSTD1k dataset，the proposed method achieves an intersection-over-union （IoU） of 67.59%，a detection rate （P_d） of 93.02%，and a false alarm rate （F_a） of 9.034×10^-6.Compared to the existing mainstream method DNANet，it improves IoU and Pd by 1.88% and 1.18%，respectively，and reduces the false alarm rate by 48.7%.This achieves a better balance between computational efficiency and detection accuracy.