Low-light wheat image enhancement using an explicit inter-channel sparse transformer

Abstract

Low-light image enhancement poses a significant challenge in agricultural settings, particularly for time-series captured wheat images. The use of time-series captured image data can improve the accuracy of wheat yield prediction by analyzing the growth status and characteristics of wheat. Given the difficulty of capturing paired wheat images in various lighting conditions, a method based on the Global Wheat Head Detection (GWHD) dataset is presented to synthesize pairs of low-light/normal-light wheat images. This approach expands to create a new dataset, LN-GWHD, specifically for the task of enhancing wheat images in low-light conditions. Importantly, it enables a more extensive evaluation of wheat image restoration tasks and performance accuracy under more controlled and diverse conditions. It is worth noting that this method can be extended to a wide range of scenarios for the enhancement of low-light situations. Where low-light conditions are common due to various factors such as weather, existing methods based on the transformer architecture have achieved state-of-the-art (SOTA) enhancement performance by capturing relationships between low-light feature channels. However, these approaches perform the similarity computation on all tokens and may introduce redundant features, limiting their ability to focus on critical information for high-quality image reconstruction. To recover more high-quality wheat image information, the explicit inter-channel sparse transformer (EIST) network is proposed, which is specifically designed to recover wheat images under low-light conditions. EIST comprises multiple blocks, each featuring explicit sparse top-k attention (ESTA) and bilateral gated feed-forward network (BGFN). ESTA dynamically selects the most critical information by focusing on the top-k key labels in each query, while BGFN enhances valid information interaction through re-double filtering. The overall architecture employs transformers to learn correlations and features between different image channels, resulting in noise reduction, detail enhancement, and improved clarity. Additionally, the Fourier spectrum loss is introduced to constrain the image reconstruction in the frequency domain, allowing more detailed information to be recovered. Extensive experiments on the LN-GWHD datasets demonstrate that EIST surpasses SOTA methods. Furthermore, our approach achieves superior results in detecting low-light wheat ears during back-end evaluation.