MIR-YOLO: Remote sensing small target detection network based on visible-infrared dual modality

The study of remote sensing detection of small targets is of great significance in the fields of traffic monitoring and military target localization and identification. However, due to small-scale targets occupying fewer pixels, they lack not only physical features such as texture and shape, but also the effective information is lost during forward propagation of the network, leading to inappropriate gradient updates, which in turn affects the accuracy of target detection. To this end, this paper introduces a Multi-order Gated Aggregation module based on Inverted Residual (MIR). This module, designed around the concept of manifolds of interest, effectively adapts to multi-scale variations and significantly mitigates information loss for small-scale targets. Furthermore, we take advantage of the complementary advantages of multi-modal information and design a multistage dual modality fusion framework, which significantly improves detection accuracy. To address the complexity and diversity of remote sensing scenes, this paper proposes a Gradient Path-Based Vision LSTM (GViL) block. This module employs the high efficiency of gradient path analysis and achieves significant results by leveraging the modeling capability of Vision LSTM (ViL) for contexts. We have verified the performance of the model in this paper on the multi-modal remote sensing datasets VEDAI and Dronevehicle, and achieved excellent results. On the VEDAI dataset, the $\text{m}A{P}_{50}$ of our model increases by 5.8% over the basic model (YOLOv8s), and by 3.7% over this year's target detection state-of-the-art method, YOLOv9.