U-Tiny: U-Net Model for Tiny Objects Detection in UAV-Images

Unmanned aerial vehicle (UAV) image recognition is particularly difficult because of the difficulty of detecting small objects. Current deep learning detectors often use the extraction of features networks with a significant down-sampling ratio to acquire coarser features. However, owing to the limitations of pixels, such a large step often causes the characteristic data of tiny objects to shift to the little spot or even disappear in the tiny feature maps. In order to gather greater characteristics of the objects with a tiny size, an innovative framework named Multi-Feature U-Net (MFU-NET) is presented in this paper. The parallel branch was developed to bring back any lost information from the more advanced levels. In the meantime, a guided spatial focus module (SSAM) is used to reduce the effect of noise inference in the background and zero in on object details. For even more localization prowess, we use the cascade structure in the Fast R-CNN step. Our approach delivers competitive performance related to existing state-of-the-art detection algorithms, as shown by research on the publicly available drone-based datasets known as VisDrone-DET.