Crested Polar Lights Optimization enabled Light Weight Relevance Capsule Network for Synthetic Aperture Radar image object classification.

To ensure continuous all-weather observation, synthetic aperture radar imagery is generated, but the large volume of data generated poses a challenge. Although DL (deep learning) models have proven to be performant at classification tasks, they usually require a high amount of storage and computational power. As a result, their deployment on mobile or edge devices is non-trivial. Moreover, the speckle noise and complicated backscattering due to the inherent SAR characteristics make accurate feature extraction difficult. In light of the above limitations, this study proposes a Crested Polar Lights Optimization-enabled Lightweight Relevance Capsule Network (CPO-LWRCapsNet) framework.As discussed, the pipeline starts by denoting the input SAR images with the help of a guided filter. Then it processes using the DHT. Next, object segmentation takes place through the Kolmogorov–Arnold Network-Driven Multiscale Synergy Network (Disney), followed by robust descriptor generation using the Relative Directional Edge Binary Patterns (RDEBP). The CPO-LWRCapsNet is a hybrid of the Crested Porcupine Optimizer (CPO) and Polar Lights Optimization (PLO). In this optimization, the RACapsNet, which is a relevance-aware capsule network, is optimized with this hybrid approach. This multi-objective optimization ensures high accuracy of classification while tuning for low parameters like memory, Flops, etc., making SAR object classification real-time yet low in complexity.