| Summary: | 碩士 === 國立中央大學 === 資訊工程學系 === 105 === In recent years, as a branch of machine learning, deep learning play an important role in Artificial Intelligence, which Convolutional Neural Network (CNN) has a breakthrough Performance in the image classification when comparing with traditional classification methods. The emergence of the full Convolutional Network (FCN)[10] also makes the study of image semantic segmentation flourish. In contrast to past work clustering according to the image texture and color, FCN joined the training of semantic information to improve the accuracy of semantic segmentation. Our paper combines the advantages of two networks, an object boundary based approach to strengthen the integrity of edge and the object itself, and the other is responsible for the prediction of image semantic segmentation, proposed an end-to-end training network architecture.
In this paper, proposed architecture improves the DT EdgeNet (Domain Transform with EdgeNet)[11]. Here, we combined the OBG-FCN [12] mask network and replaced the [11] edge network. The used mask network can predict background, object, and object edge reference diagrams. In addition, our architecture uses multi-scale ResNet-101 as the base network and introduces multi-scale Atrous Convolution to architecture training to preserve the dimensions of the feature map, which increases the receptive and further to enhance the accuracy of semantic segmentation.
In the experiments, we got the high performance of recognition on the VOC2012 test set. In addition, we combined extraction of object bounding box generated by Faster RCNN and result of proposed semantic segmentation as an extension application for instance-level segmentation.
|
|---|
