基于特征提取的RGBD语义分割算法研究

Research on RGBD Semantic Segmentation Algorithm Based on Feature Extraction

摘要: RGBD语义分割是近年来备受关注的研究领域。该领域的挑战在于有效地利用RGB和深度图像的不同信息特征。RGB图像具有全局颜色变化的特点，而深度图像则提供了关于对象局部位置的信息。因此，深度图像被认为是更具代表性的局部语义信息源，有助于后续的编码和分割。然而，目前的方法往往将RGB和深度图像通过相同的卷积运算符进行处理，忽略了它们之间的固有差异。为了解决这一问题，本文对传统的重叠补丁嵌入方法进行了修改，以更好地利用深度信息，实现更高精度的语义分割。通过修改传统的重叠补丁嵌入方法，本文能够更好地利用深度信息。具体来说，本文提出了一种改进的方法，通过对深度图像进行处理，更好的进行特征提取。通过在数据集上进行实验，本文的方法在RGBD语义分割任务中取得了较高的准确性和鲁棒性。

Abstract: RGBD semantic segmentation is a research area that has received much attention in recent years. The challenge in this area is to effectively utilize the different information characteristics of RGB and depth images.RGB images feature global color changes, while depth images provide infor-mation about the local location of an object. As a result, depth images are considered to be a more representative source of local semantic information, which helps in subsequent coding and seg-mentation. However, current approaches tend to process RGB and depth images by the same con-volution operator, ignoring the inherent differences between them. To address this problem, this paper modifies the traditional overlapping patch embedding method to better utilize the depth in-formation and achieve higher precision semantic segmentation. By modifying the traditional over-lapping patch embedding method, this paper is able to better utilize the depth information. Specif-ically, this paper proposes an improved method for better feature extraction by processing depth images. By conducting experiments on the dataset, the method in this paper achieves high accuracy and robustness in RGBD semantic segmentation tasks.