高效并转融全景玻璃精分网络

Efficient and Integrated Panoramic Glass Precision Sorting Network

摘要: 精准分割环境中的玻璃物体，是提升自动驾驶、深度感知等视觉系统性能的关键环节。然而当前主流的深度学习分割方法，其训练与推理几乎完全建立在传统的透视图像之上，这类图像的有限视野与局部的上下文信息，使其在处理开阔场景中尺度多变、距离各异的玻璃目标时显得力不从心。虽然全景成像能提供无死角的全局环境感知，但其中玻璃区域因透视产生的剧烈形变，与其自身的透光、反射等固有光学特性相互作用，构成了一个极度复杂的视觉分析难题，远超传统透视图像所面临的挑战。为系统性地解决上述难题，本文提出了一种新的网络架构——高效并转融全景玻璃精分网络。该神经网络架构系统集成了注意力机制、转置卷积、深度可分离卷积和空卷积等先进操作，分别设计了三个模块：高效并行卷融深度可分模块、高效转卷双支融调模块和高效并转累融精调模块，用于对主干网络提取的特征进行再处理。我们在PanoGlass V2等基准数据集上的实验表明，本方法关键指标显著优于现有技术，IoU、MAE与F-Score分别达到91.37%、95.49%与0.0060，验证了其高效性与优越的泛化能力，为复杂场景下的全景视觉应用提供了可靠解决方案。

Abstract: Accurately segmenting glass objects in the environment is a crucial step in enhancing the performance of visual systems such as autonomous driving and deep perception. However, the current mainstream deep learning segmentation methods rely almost entirely on traditional perspective images for training and inference. The limited field of view and local contextual information of such images make them inadequate in handling glass targets with varying scales and distances in open scenes. Although panoramic imaging provides a comprehensive and unobstructed view of the environment, the severe deformation of glass areas due to perspective, coupled with their inherent optical properties such as light transmission and reflection, poses an extremely complex visual analysis challenge, far exceeding the challenges faced by traditional perspective images. To systematically address the aforementioned challenges, this paper proposes a novel network architecture—Efficient and Integrated Panoramic Glass Precision Sorting Network. This neural network architecture integrates advanced operations such as attention mechanisms, transposed convolution, depthwise separable convolution, and spatial convolution, and designs three modules: the Efficient Parallel-to-Global Deepwise Separable Module, the Efficient Transposed-to-Global Dual-Stream Fusion Module, and the Efficient Parallel-to-Global Accumulative Fusion Fine-tuning Module, for reprocessing the features extracted by the backbone network. Our experiments on benchmark datasets such as PanoGlass V2 demonstrate that the key metrics of this method significantly outperform existing techniques, achieving IoU, MAE, and F-Score of 91.37%, 95.49%, and 0.0060, respectively. This verifies its efficiency and superior generalization ability, providing a reliable solution for panoramic vision applications in complex scenes.