A deep neural network based feature fusion panoramic segmentation method

Abstract Segmentation of complex and diverse scenes in the field machine vision has advantages low cost good robustness. However, real open intelligent driving scene, segmented object will be affected by change proportion illumination. The feature extraction interference algorithm causes some problems such as local information loss background fusion process execution. In this paper, deep neural network method is used for panoramic segmentation. It focuses on premise lightweight, improves performance improving structure semantic segmentation layer model calculation effect, reducing number parameters calculations, so to further improve detection performance. At same time, with gradual deepening learning architecture, but most embedded devices can not bear a huge amount computing workload. An attentional channel pruning based approach proposed using wide range experimental ablation studies compare different data, including VOC2012 ImageNet datasets. We show that outperforms state art methods terms visual quality.