基于全局卷积神经网络的复杂图像语义分割方法

A Method for Semantic Segmentation of Complex Images Based on Global Convolutional Neural Network

语义分割的场景图像易受不同光照强度以及类别多样性的影响,尤其是在复杂的图像分割任务中,由于不同物体间的像素值差异过大或过小,造成分割图像的纹理和几何特征缺失,即产生欠分割、过分割现象。针对上述问题,利用深度卷积神经网络,研究基于全局卷积神经网络的复杂图像语义分割方法。首先,提出多尺度残差空间金字塔池化模块,在网络中获取到更加稠密和完备的图像低层特征[1];其次,网络考虑全局信息,提出基于注意力机制的解码器模块,有效捕获图像像素的纹理特征、颜色特征和上下文信息,从而得到完整的分割结果。该方法在Camvid数据集上分割精确度达68.5%(MIoU)且在Cityscapes数据集上分割精度达78.3%。

Semantic segmented scene images are susceptible to different light intensities and category diversity,especially in complex street scene segmentation tasks,because the pixel values between different objects are too large or too small,resulting in the lack of texture and geometric features of the segmented image.That is under-segmentation and over-segmentation.Aiming at the above problems,a deep convolutional neural network is used to study the semantic segmentation of complex streetscape images based on the global convolutional neural network.First,a multi-scale residual spatial pyramid pooling module is proposed to obtain more dense and complete low-level image features in the network.Second,the network considers global information and a decoder module based on the attention mechanism is proposed to effectively capture the texture of image pixels.Features,color features and contextual information are obtained to get a complete segmentation result.This method has a segmentation accuracy of 68.5%(MIoU)on the Camvid dataset and has a segmentation accuracy of 78.3%on the Cityscapes dataset.