A Semantic-based Scene segmentation using convolutional neural networks

摘要

Semantic segmentation is a crucial operation in the computer vision field. One of the promising techniques is the convolutional neural network (CNN). It can be utilized with both single and multidimensional arrays and is useful for processing 2D arrays in computer vision tasks. In this paper, a new model for semantic scene segmentation is proposed. In order to enhance the segmentation results, the model starts with classifying the input scene as either indoor or outdoor scenes. In this context, the MobileNet is used as it provides better results when compared to Inception-v3 and Inception-ResNet-v2 networks. The next step, two models based on Pyramid Scene Parsing Network (PSPNet) are used for image segmentation (indoor images are segmented by the indoor model and outdoor images are segmented by the outdoor model). Experimental results prove the concept that a specific scene model can achieve higher accuracy than general scene models on the semantic segmentation task. (C) 2020 Published by Elsevier GmbH.