An Enhanced Multi-Objective-Derived Adaptive DeepLabv3 Using G-RDA for Semantic Segmentation of Aerial Images

Semantic segmentation acts as a major role in classifying the remote sensing images into oceanic ice, vegetation, roads, vehicles, houses, and more for offering high precision at the pixel level. In recent studies, convolutional neural network (CNN) has accomplished superior efficiency in the semantic segmentation of images. Even though various deep techniques and architectures have been utilized for enhancing the accuracy, it suffers from classifying the confused classes. Due to the optical conditions and remote sensing information, the sub-decimeter aerial imagery segmentation is challenging while achieving fine-grained semantic segmentation outcomes. The core goal of this task is to adopt the latest Adaptive DeepLabv3 strategy for enhanced semantic segmentation of aerial images. In Adaptive Deeplabv3, the involvement of both the encoder-decoder structure and spatial pyramid pooling module with adaptiveness by a hybrid meta-heuristic algorithm makes faster and stronger segmentation performance within less search space and reduced computation time. The relevant parameters of DeepLabv3 are tuned or optimized by the hybrid meta-heuristic algorithm based on Genetic Inspired Red Deer Algorithm (G-RDA). The enhanced segmentation is employed concerning a fitness function with precision and accuracy. Finally, the experimental analysis of the suggested Adaptive DeepLabv3 strategy for semantic segmentation of aerial images proves its competitive solution when evaluated over conventional approaches.