Deep Learning-Based UAV Image Semantic Segmentation Algorithm Research

Abstract

Most existing image semantic segmentation algorithms for UAV vision are limited to remote sensing images, which lack the resolution to accurately represent ground details, thereby hindering UAV's real-time autonomous environment perception in low-altitude flight missions.To address this issue, a real-time image semantic segmentation method for low-altitude UAV is proposed.A new hyper-network architecture is designed.A context header weight generation module is added to the last layer of the encoder, and the weight of each block in the decoder is generated before the end of the encoder encoding, to reduce the number of network parameters and computation during prediction and achieve the effect of real-time segmentation.In the decoder, a dynamic fragment convolution algorithm is designed using the local connection layer mechanism.When facing large segmented objects that span multiple fragments, the semantic information of the context is fully considered, to ensure that the weight of each convolution core in the decoder changes with the spatial position of the input feature map.Simultaneously, the dynamic weight is used to segment different objects in a targeted manner, maximizing the adaptability of the network.The experimental results on the low altitude UAV vision image dataset demonstrate that the mean Intersection over Union(mIoU) of this method for buildings, roads, static vehicles, and other categories is 66.3%, and the prediction speed reaches 37.9 frame/s.Compared with MSD and ABCNet algorithms, its segmentation accuracy improved by 9.3 and 2.5 percentage points, respectively.