作者:禅与计算机程序设计艺术
1.背景介绍
Semantic image segmentation (SIs) is one of the most challenging tasks in computer vision and medical imaging fields due to its high variability and complexity of realistic scenes with complex structures and textures. In this paper, we present an attention-based fully convolutional network (ABFCN) architecture that leverages a deep learning technique called conditional random field (CRF) post-processing to segment semantic regions accurately from RGB images. The proposed approach first uses a feature extractor to extract feature maps at multiple scales for both foreground and background objects using resnet50 as backbone. Then, we apply an attention mechanism to selectively focus on important features at different scales. Next, we use four parallel blocks of ABFCNs each consisting of two branches of convolution layers followed by batch normalization and ReLU nonlinearity. The output of these blocks are concatenated along