High Dimensional Spatial Information and Multi-scale Fusion Network for Efficient and Real-Time Small Object Detection in Remote Sensing Images

Abstract

The detection of small objects in remote sensing imagery remains a formidable challenge due to their minimal pixel occupancy, blurred structural boundaries, and susceptibility to environmental interference. To solve these problems, this paper proposes a novel network architecture named multidimensional information feature fusion-you only look once (MIFF-YOLO), which integrates several specialized modules. To address the challenge of small objects being obscured by complex environmental factors, we propose a multidimensional information fusion (MIF) module for the neck network, which leverages a 3D convolution and a full-domain transformer (FDT) to create cross scale dependencies and integrate global contextual information with local details. For the purpose of preserving the spatial and edge information of small objects, an efficient front end module (EFEM) is embedded into the C3k2 architecture. The EFEM module employs a parallel, learnable dual-path architecture that collaboratively integrates a Sobel convolution stream for explicit edge detection and a spatial information stream max-pooling for detail preservation, enabling simultaneous extraction of structural boundaries and contextual textures. These complementary features undergo an adaptive fusion via omni-dimensional dynamic convolution (ODConv), thereby enriching the capabilities of the feature representation. In order to address the loss of critical details in small object features during enlargement, dynamic upconvolution block (DUB) is introduced to replace standard upsampling module. Adaptive feature sampling is achieved through content-aware dynamic offsets, mitigating detail loss during resolution recovery. Compared with the original baseline algorithm, the improved network achieved a 3.7% improvement on mAP@50 and a 3.9% improvement on mAP@50:95, with the FPS reaching 120 on the DOTA dataset. This shows that the improved algorithm effectively enhances small object detection performance in remote sensing images while maintaining excellent real-time detection efficiency.