SqueezeSlimU-Net: An Adaptive and Efficient Segmentation Architecture for Real-Time UAV Weed Detection

Abstract

The limited processing capacity of computing equipment that is usually mounted on unmanned aerial vehicles (UAVs) often prevents real-time execution of computer vision tasks, such as image segmentation. In this article, we introduce SqueezeSlimU-Net (SSU-Net), an adaptive and efficient deep learning (DL) model designed to enhance UAV capabilities in performing complex image segmentation tasks under resource constraints, thereby advancing real-time UAV vision - a crucial technology in fields, such as precision agriculture. SSU-Net combines benefits of three specialized DL architectures: the semantic segmentation capabilities of the U-Net architecture, the computational efficiency of SqueezeNet's fire modules, and the dynamic adaptability of slimmable neural networks. This integration allows SSU-Net to adjust its network width in real-time, thus striking the balance between inference accuracy and computational load based on the operational parameters such as task requirements and UAV's battery life. To validate SSU-Net's efficacy, we applied it to a weed detection task using two UAV-collected datasets and tested it on an edge computing platform for UAVs. Our experiments show that SSU-Net can reduce inference energy consumption by up to 65% with only a minimal 2% reduction in accuracy. A comparative evaluation with other state-of-the-art DL image segmentation approaches shows that SSU-Net achieves on par weed detection performance while requiring significantly fewer model parameters. In addition, SSU-Net outperforms state-of-the-art network pruning techniques in balancing accuracy and resource usage. Timing benchmarks show SSU-Net fostering real-time weed detection even on low-resource UAVs, making it ideal for UAV remote sensing applications.