Prune and Quantize Semantic Segmentation Network for Aerial Objects Recognition

Abstract

Semantic segmentation of aerial and satellite images is crucial for applications in environmental management, urban planning, and traffic safety. While deep learning techniques with convolutional neural networks (CNNs) and attention mechanisms have achieved superior accuracy compared to traditional methods, they often struggle with model complexity and resource constraints. This paper introduces two novel techniques - pruning and quantization - to enhance both the performance and efficiency of semantic segmentation models for remote sensing images (RSIs). Pruning reduces model complexity by eliminating less significant weights, while quantization decreases memory usage by converting weights into a more compact format. We applied these techniques to the DeepLabV3+ model with ResNet18 and ResNet50 backbones and assessed their performance across multiple RSI datasets. Our results show that pruning and quantization effectively balance accuracy and computational efficiency, achieving a mean IoU of 81.24% with a memory footprint of 135.19 MB for pruning, and 81.04% mean IoU with a memory footprint of 33.79 MB for quantization on the ISPRS Vaihingen dataset. These methods offer a viable solution for deploying semantic segmentation models on resource-constrained hardware.