MIDFE: Multi-image differential feature enhancement based on misalignment

Single-camera systems face the challenge in balancing wide field of view (FOV) and high resolution. Multi-camera systems, including camera arrays, offer an efficient approach to overcome this difficulty and have been widely adopted in large-scale surveillance and remote sensing domains. A multi-image differential feature enhancement (MIDFE) network designed for array camera systems is proposed in this paper. The differential feature enhancement (DFE) module utilizes differential mapping and attention mechanisms to enhance features. We conclude that the minor misalignments introduced during the image alignment stage can facilitate more effective fusion of multi-view features within the network, leading to improved detection performance. Additionally to accommodate the needs of our model, we converted the single-view DOTA dataset into simulated array camera perspectives. The experimental results show that the proposed module improves the performance of the target detection systems while maintaining efficiency.