Self-calibrated acceleration and detail preserving for semantic segmentation of lactating sows and piglets under low-light conditions

Abstract

Abstract Pixel-wise semantic segmentation of lactating sows and piglets is critical to explore maternal traits in smart livestock breeding and production. However, semantic segmentation algorithms do not yield satisfactory results under low-light conditions because these methods are highly dependent on image quality. Therefore, an efficient and detail-preserving low-light image enhancement is necessary and crucial for animal monitoring under low-light conditions. In this paper, a low-light enhancement method called self-calibrated acceleration and detail preserving (SADP) was proposed to improve the semantic segmentation performance of lactating sows and piglets. Specifically, a self-calibration acceleration module that accelerated the convergence among all stages was proposed to improve the computational efficiency and a semantic perceptual loss term was proposed for a high detail and semantic information preservation. Plenty of experiments demonstrated that SADP outperformed the existing well-known methods in both visual quality and efficiency (-0.013s in execution time from 0.0163s to 0.0033s, -0.0031 M in mode size from 0.0034 M to 0.0003 M), and even more improved the performance of semantic segmentation of lactating sows and piglets, raising the mean IOU from 0.8686 to 0.8872. Obviously, SADP also can be used to improve the performance of other high-level visual tasks. This may build a good foundation for the following visual tasks and further promote the livestock breeding and production.