Dim-Light Instance Segmentation Based on Soft Thresholding and Improved BCNet

Instance segmentation is a new method in the field of security monitoring. The current models perform well in sufficient light. However, in the actual environment, due to the change of light in the scene, the picture will randomly appear with high noise, low brightness and other problems which seriously affect the segmentation accuracy. Therefore, two S-BCNet based on BCNet are proposed in this paper, which respectively achieve better performance under complex light and constant dim-light conditions. In order to improve the generalization, we propose a new Shrinkage Module based on a soft thresholding function and insert it into BCNet's backbone so that it can learn and filter noise autonomously. Moreover, we propose a dim-light processing method and produce two datasets: COCO-D(dark light) and COCO-M(complex light) based on COCO, which are specifically used to train and test the performance of the mode in various degrees of dim-light environments. Furthermore, we use AdaBN to improve our S-BCNet and use the COCO-D for transfer learning to train the model for dark-light specialization, specifically for use in constant dim-light scenes. Experiments show that our models improve by 10.7% and 17.2% respectively compared with the best mainstream model in the two environments.