加强空间信息引导的道路场景实时语义分割

Real-time semantic segmentation of road scenes with enhanced spatial information guidance

针对实际道路场景不同类别位置和轮廓等细节部分相差较大,导致细节信息的丢失和小目标分割不准确等问题,提出一种加强空间信息引导的实时语义分割网络。该网络首先采用微调的轻量级分类卷积编码器,提取图像中不同层次的高级语义信息和浅层空间信息;其次,设计空间细节引导融合模块为编码器的深层与浅层的特征映射提供指导,使得融合过程中更加信任细节轮廓信息,增强空间感知能力,并抑制背景噪声影响;最后通过损失函数辅助监督训练和数据相关型上采样加强训练阶段的特征表示,进一步优化深度卷积网络,以弥补大幅度上采样造成的信息损失。在CamVid和Cityscapes数据集上进行实验验证,其分割精度分别为73.7%和75.3%,推理速度分别为158.0和126.5 fps。与其他实时语义分割算法相比,算法能更好地平衡精确度和实时性,在实际应用场景中也具有更好的鲁棒性。

Actual road scenes vary greatly in different categories of locations and contours and other detailed parts,which can lead to problems such as loss of detailed information and inaccurate segmentation of small targets.A real-time semantic segmentation network with enhanced spatial information guidance was proposed.Network used a fine-tuned lightweight classification convolutional encoder to extract high-level semantic information and shallow spatial information at different levels in the image.Furthermore,design of the spatial detail-guided fusion module provides guidance for deep and shallow feature mapping of encoder,enabling greater trust in detailed contour information during fusion,enhancing spatial perception and suppressing the effects of background noise.Finally,feature representation in training phase is enhanced by loss function assisted supervised training and data-dependent upsampling to further optimize deep convolutional network to compensate for the loss of information due to significant upsampling.Experiments on CamVid and Cityscapes datasets show that accuracy of this method is 73.7%and 75.3%,respectively,and inference speed is 158.0 and 126.5 fps,respectively.Compared with other real-time semantic segmentation algorithms,it can achieve a better balance between accuracy and real-time,and has better robustness in practical application scenarios.