Low Illumination Instance Segmentation Based On RGB-D Images

The instance segmentation in low-illumination scenes poses a challenge to the robotic understanding of the environment due to the influence of ambient illumination,and the limitation of depth information itself also leads to less than ideal results of segmentation.Most of the existing instance segmentation algorithms focus on the improvement of accuracy,which leads to the growth of model parameters and reduces the processing speed.Therefore,it is important to carry out research on low-illumination instance segmentation based on RGB-D images to improve the accuracy and network lightweight of instance segmentation under low illumination and expand the application scenarios of instance segmentation.Aiming at the above problems,instance segmentation in low illumination environment was carried out based on multi-scale Retinex algorithm and Depth information and YOLACT network,and the major research works are as follows.(1)RGB-D instance segmentation dataset and low-lighting dataset in indoor office environment.Images are acquired with Realsense d435 i and labeled with labelme to construct the RGB-D instance segmentation dataset in COCO format.The images are processed by inverse gamma transform to obtain the low-illumination dataset.(2)A low-light image enhancement algorithm based on improved multi-scale Retinex.Aiming at the problems of traditional Retinex algorithm such as poor color retention and detail loss,this paper constructs the main feature layer-compensation layer structure and combines the artificial bee colony algorithm to optimize the fusion weights to obtain a low-illumination image enhancement algorithm.(3)YOLACT instance segmentation network based on RGB-D images.In order to lighten the instance segmentation network and improve the segmentation efficiency,the paper combines Bi FPN and multi-information fusion modules to carry out the research of target instance segmentation in indoor office scenes and improve the YOLACT instance segmentation network under RGB;Meanwhile,for the problem of unsatisfactory instance segmentation in low-light environment,the paper incorporates Depth features and image enhancement algorithms to carry out the research of instance segmentation in low-light and improve the RGB-D YOLACT instance segmentation network.The experiment results show that the improved multi-scale Retinex enhancement algorithm,which improves the enhanced images in Information Entropy(IE),Average Gradient(AG)and Standard Deviation(SD)by 84.4%,1044.7% and 1065.7%,respectively,and performs well in subjective evaluation.The improved YOLACT instance segmentation network improves the FPS by 22.9% and has significant improvement in m AP compared with the traditional YOLACT network.Also,after incorporating Depth information and combining with image enhancement algorithm,the target objects can be marked and segmented more efficiently in low illumination environment.