Image Super-Resolution for Arthropod Identification

Duhita Wani, Tomás Maul

Abstract: Super-resolution techniques have recently made great strides, especially in the context of deep learning. In spite of this, not much research has been conducted on the explicit application of these techniques to biodiversity related problems such as species identification. We took a state-of-the-art super-resolution model (enhanced deep super-resolution network, i.e. EDSR), and enhanced it further with perceptual and texture losses, and a test-time-augmentation solution. Furthermore, we designed a qualitative assessment framework and studied its relationship with automated performance metrics. Our results show that our proposed modifications to EDSR improve the recovery of details, and that current automated metrics (e.g. Peak Signal-to-Noise Ratio) are inadequate in the context of super-resolution for species identification.