基于自适应注意力机制的轻量化语义分割网络

Lightweight Semantic Segmentation Network Based On Adaptive Attention Mechanism

摘要:

针对语义SLAM(simultaneous localization and mapping)中语义分割速度较慢,实时性较低、占用资源过多等问题,提出一种含有自适应通道注意力机制的轻量级Mask R-CNN网络,由于原有的语义分割网络里的残差网络复杂,且应用环境在室内,环境较为简单,故该轻量级网络将原有复杂的主干网络中的ResNet-50利用深度可分离卷积与分组卷积改进为更加轻量的ResNet-DS-tiny(ResNet with depthwise separable convolutions),并加入自适应通道注意力机制。在自适应通道注意力模块中,利用加权方式对输入的RGB-D图像从空间和通道赋予不同的权重,增强了特征的表达能力。此外,为了轻量化特征金字塔,使用使用不同空洞率的空洞卷积来提取不同大小感受野的特征信息,有效地获取了多尺度的特征。相较于传统的特征金字塔,空洞卷积减少了参数量。在更充分获取 RGB 信息特征的同时,提升了语义分割系统的实时性并减少了资源占用。

Abstract:

To address the issues of slow semantic segmentation speed, low real-time performance, and high resource consumption in semantic SLAM (simultaneous localization and mapping), a lightweight Mask R-CNN network with an adaptive channel attention mechanism is proposed. Given the complexity of the residual networks in existing semantic segmentation networks and the relatively simple indoor application environments, this lightweight network replaces the original complex backbone ResNet-50 with a more lightweight ResNet-DS-tiny (ResNet with depthwise separable convolutions) by incorporating depthwise separable convolutions and grouped convolutions. An adaptive channel attention mechanism is also introduced. In the adaptive channel attention module, a weighted approach is used to assign different weights to the input RGB-D images from both spatial and channel dimensions, thereby enhancing the feature representation capability. Additionally, to lighten the feature pyramid, dilated convolutions are employed to expand the receptive field, effectively aggregating multi-scale features with different dilation rates. Compared to traditional feature pyramids, the use of dilated convolutions reduces the number of parameters. This approach not only more effectively captures RGB information features but also improves the real-time performance of the semantic segmentation system while reducing resource consumption.