Feature enhanced spherical transformer for spherical image compression

Abstract

It is well known that the wide field of view of spherical images requires high resolution, which increases the challenges of storage and transmission. Recently, a spherical learning-based image compression method called OSLO has been proposed, which leverages HEALPix’s approximately uniform spherical sampling. However, HEALPix sampling can only utilize a fixed 3 × 3 convolution kernel, resulting in a limited receptive field and an inability to capture non-local information. This limitation hinders redundancy removal during the transform and texture synthesis during the inverse transform. To address this issue, we propose a feature-enhanced spherical Transformer-based image compression method that leverages HEALPix’s hierarchical structure. Specifically, to reduce the computational complexity of the Transformer’s attention mechanism, we divide the sphere into multiple windows using HEALPix’s hierarchical structure and compute attention within these spherical windows. Since there is no communication between adjacent windows, we introduce spherical convolution to aggregate information from neighboring windows based on their local correlation. Additionally, to enhance the representational ability of features, we incorporate an inverted residual bottleneck module for feature embedding and a feedforward neural network. Experimental results demonstrate that our method outperforms OSLO, achieving lower codec time.