Abstract: Accurate and efficient military aircraft target detection algorithms can significantly enhance the perceptual capability of surveillance systems. To address the issues of missed detections, false alarms, and difficulty in meeting real-time monitoring requirements for military aircraft in remote sensing imagery, this paper proposes a lightweight YOLO-based remote sensing military aircraft detection method, Slim-YOLO. First, a lightweight dual-head small-object detection network is designed. By optimizing the receptive field for small targets and fusing multi-scale features, it alleviates the loss of fine-grained details during convolution. Second, a Feature Median Redundant Filter Pruning algorithm (FMRFP) is introduced, which measures the replaceability of convolution kernels via the feature median and removes redundant filters, achieving highly sparse compression while preserving scattered details. Finally, a CVSIoU (Cosine-Varifocal Shape Intersection-over-Union) loss is constructed. It incorporates target size–adaptive adjustment and an angle-balancing mechanism to reweight positive samples and optimize overlap-region loss, thereby reducing the negative impact of redundant anchors on detection accuracy. Experiments on the MAR20 dataset show that Slim-YOLO improves detection accuracy to 91.75%; after pruning, the model size is compressed to 190 KB and inference time is reduced to within 3 ms. Compared with mainstream methods, Slim-YOLO achieves superior performance in accuracy, model size, and inference speed, meeting the real-time detection requirements for military aircraft in remote sensing imagery.