A Lightweight Normalization-Free Architecture for Object Detection in High-Spatial-Resolution Remote Sensing Imagery

High-spatial-resolution remote sensing imagery applications demand computationally efficient detection methods due to their massive scale and real-time processing requirements. However, fast and reliable object detection in high-spatial-resolution remote sensing imagery is very challenging. Global channel-attention modules are computationally intensive, and batch-normalization layers incur extra computation, increasing latency for megapixel-scale inputs. To address these limitations, we propose Lino-YOLO, a lightweight, normalization-free architecture specifically designed for resource-constrained remote sensing environments. With only 2.64 million parameters, Lino-YOLO integrates three efficiency-oriented components. These include coordinate attention, which enables precise spatial context extraction at minimal computational cost; dynamic tanh, which eliminates the need for batch normalization and provides lightweight training stability; and wise-IoU, which optimizes gradient allocation without additional computational overhead. Our lightweight design achieves remarkable efficiency, demonstrated by a 7.6 ms inference time on an RTX 4090 GPU, while delivering superior accuracy—93.7% [email protected] on the NWPU-VHR10 benchmark. Comprehensive evaluation across multiple YOLO families (v8–v11) demonstrates consistent improvements, with gains ranging from 6.6 to 11.0 percentage points over their respective baselines. Further evaluation on the challenging DIOR dataset reveals consistent improvements across all categories, confirming the effectiveness and robustness of our approach. This exceptional performance-to-size ratio makes Lino-YOLO ideal for large-scale, real-time remote sensing tasks, where both speed and accuracy are critical.