Research On Semantic Segmentation Algorithm Of UAV Image Based On Deep Learning

The autonomous environmental perception technology of unmanned aerial vehicles has the ability to quickly monitor and analyze complex environments at specific locations and times,providing necessary technical support and data foundation for applications such as traffic monitoring,population density monitoring,and urban greening monitoring.Existing research mainly focuses on target detection and large-scale house segmentation tasks under drone vision,while there is still a lack of good solutions for real-time image semantic segmentation tasks under low altitude drone vision.In addition,in order to improve accuracy,current research on semantic segmentation of unmanned aerial vehicle visual images often adopts complex redundant models,but there are still challenges in achieving both "high accuracy" and "low latency".In response to the above issues,this article conducts research on real-time image semantic segmentation technology for unmanned aerial vehicle vision,mainly focusing on:(1)This paper first aims to verify the feasibility of using lightweight image semantic segmentation methods in unmanned aerial vehicle(UAV)visual image semantic segmentation tasks,and to explore their advantages and disadvantages in this field.ICNet,ESPNet,and DFANet,three lightweight image semantic segmentation methods,are studied in detail,and their architectural differences and respective strengths and weaknesses are analyzed.Then,the selected lightweight image semantic segmentation network is applied to UAV visual image semantic segmentation tasks,and corresponding experiments are designed to verify the effectiveness of the algorithm.Finally,this paper deeply analyzes the advantages and disadvantages of existing lightweight image semantic segmentation algorithms in UAV visual image semantic segmentation tasks through interpretation of experimental results and visualization of segmentation effects,providing theoretical basis and targeted improvement directions for future work.(2)To address the issues mentioned earlier and analyze the characteristics of drone visual images in depth,this paper proposes an innovative semantic segmentation method for low altitude drone real-time images.The method adopts an efficient hypernetwork architecture and generates the weights of each block in the decoder in real-time through a context header weight generation module introduced in the last layer of the encoder,achieving real-time semantic segmentation.A dynamic sliced convolution algorithm is designed in the decoder using a local connectivity layer mechanism,which fully considers the semantic information of the context,and performs targeted segmentation for different objects to improve the adaptability of the network.To solve the problem of small datasets in UAV visual image semantic segmentation,transfer learning is used to pretrain the proposed network on datasets with similar image features,allowing the network to obtain better training performance and more advanced feature information.The parameter tuning and ablation experiments are conducted on the proposed model,and the model is compared and quantitatively evaluated.The proposed model shows the excellent results,with an average intersection-over-union of66.3% for each category and a prediction speed of 37.9 frames/s.The segmentation accuracy is significantly improved while ensuring real-time performance,demonstrating the superiority and applicability of the proposed algorithm.The algorithm in this article has been proven through experiments to accurately and efficiently complete the semantic segmentation task of unmanned aerial vehicle visual images,which has high research significance and practical value.