Semantic segmentation for multiscale target based on object recognition using the improved Faster-RCNN model

Image semantic segmentation has received great attention in computer vision, whose aim is to segment different objects and provide them different semantic category labels so that the computer can fully obtain the semantic information of the scene. However, the current research mainly focuses on color image data as training, for outdoor scenes and single task semantic segmentation. This paper carries out multi-task semantic segmentation model in the complex indoor environment on joint target detection using RGB-D image information based on the improved Faster-RCNN algorithm, which can simultaneously realize the indoor scene semantic segmentation, target classification and detection multiple visual tasks. In which, in view of the influence of uneven lighting in the environment, the method of fusion of RGB images and depth images is improved. While enhancing the fusion image feature information, it also improves the efficiency of model training. Simultaneously, in order to meet the needs for operating on multi-scale target objects, the non-maximum value suppression algorithm is improved to improve the model's performance. So as to realize the output of the model's multi-task information, the loss function has also been redesigned and optimized. The indoor scene semantic segmentation model constructed in this paper not only has good performance and high efficiency, but also can segment the contours of different scale objects clearly and adapt to the indoor uneven lighting environment. (C) 2021 Published by Elsevier B.V.

Automated summary

This paper proposes a multi-task semantic segmentation model in complex indoor environments using the improved Faster-RCNN algorithm, addressing issues of uneven lighting by enhancing fusion methods and algorithms. The model achieves high performance and efficiency in segmenting contours of different scale objects.