Original scientific paper

https://doi.org/10.17559/TV-20231023001051

Advancing UAV Image Semantic Segmentation with an Improved Multiscale Diffusion Model

Wang Wang ; Geely University of China, Chengdu, Sichuan, 641423, China *
Chen Zhou ; Wuhan University, Wuhan, Hubei, 430072, China
Hua He ; Chongqing Technology and Business University, Chongqing, 400067, China
Changsong Ma ; Krirk University, Bangkok, 10220, Thailand

* Corresponding author.

Abstract

This study explores the challenges of image semantic segmentation in autonomous driving across varied campus environments. We introduce a specialized dataset consisting of 400 drone-captured images from different campuses. These images have been meticulously labelled into five categories: buildings, vegetation, ground, playgrounds, and lakes. These categories are essential for precise semantic segmentation tasks which are crucial to autonomous driving applications. To address the segmentation challenges presented by the unique and diverse features of campus environments, we propose an innovative algorithm. This algorithm is based on an enhanced diffusion model that is adept at handling multi-scale features inherent in campus environments. By incorporating scalable jump-connection layers in the denoising probability diffusion model, the proposed algorithm not only achieves superior accuracy but also demonstrates a significant improvement in recognition precision within the dataset, resulting in an average mIoU of 85%. The results underscore the algorithm's effectiveness and its potential as a robust solution for semantic segmentation tasks in autonomous driving within campus settings, paving the way for further research and application in real-world scenarios.

Keywords

image semantic segmentation; multiscale diffusion model style; reviewing; UVA

Hrčak ID:

321906

URI

https://hrcak.srce.hr/321906

Publication date:

31.10.2024.

Visits: 9 *