结合分数阶边缘检测的改进DeepLabv3+模型

Improved DeepLabv3+Model Combined with Fractional Edge Detection

DeepLabv3+是目前语义分割领域最为先进的模型之一,Google将DeepLabv3模型作为编码器的基础上创新性的加入解码器结构,强化了模型恢复空间分辨率与位置信息的能力,显著提升了分割效果,并在PASCAL VOC2012数据集上达到了SOTA(State Of The Art)。但模型仍然保留在编码器中引入下采样和空洞卷积进行特征提取,造成大量细节信息丢失,最终导致图像的边缘分割效果不佳。本文提出一种Deep Labv3+结合分数阶(微分)边缘检测的语义分割模型,采用分数阶滑动阈值边缘检测模块来强化目标图像的边缘信息,经过模型运算后得到边缘细节改善的分割结果。实验结果表明,该方法在图像的边缘分割效果上优于原始DeepLabv3+方法,并且在PASCAL VOC2012数据集上达到了89%以上MIOU的前沿水平。

DeepLabv3 + proposed by Google Inc. is one of the advanced semantic segmentation model,which creatively added the decoder structure with the previous version as an encoder part to enhance the ability of the model to recover spatial resolution and location information,scored the state-of-the-art MIOU on the PASCAL VOC2012 dataset. However,the model does not perform well in achieving good results on the image edge segmentation,due to the structures of the down-sampling and spatial convolution,lots of detail information were lost in the training process. In this paper,a semantic segmentation method that combined the DeepLabv3 + model with a fractional-order differential edge detection module is proposed,and the slip threshold method of the fractional-order differential operator had been used to enhance the edge information of the images. Our method in this paper shows better performance than the original DeepLabv3 + model on the image edge segmentation,and also scored over 89% MIOU on the PASCAL VOC2012 dataset.