Optimizing Lightweight Convolutional Neural Networks for Hyperspectral Image Fusion

Abstract

This paper addresses the challenges associated with processing complex scenes and high-precision images on resource-constrained devices, particularly within the domain of hyperspectral image processing and super-resolution reconstruction techniques. We present an optimized model for existing hyperspectral image fusion models, leveraging network lightweight and channel pruning. Our proposed model, NestFuse Small, employs NestFuse as the primary network architecture and integrates a quantization pruning module. Experimental findings demonstrate that, in comparison to the original NestFuse model, NestFuse Small’s computation speed is 164.8% of the pre-optimization speed, a decrease in memory usage of 20.65%, resulting in a slight decrease in performance. This optimized model facilitates more efficient image processing on resource-constrained devices.