JP Journal of Heat and Mass Transfer
Special Issue II, Advances in Mechanical System and ICT-Convergence, Pages 1 - 10
(December 2020) http://dx.doi.org/10.17654/HMSIII20001 |
|
ATTENTION-BASED SALIENT OBJECT DETECTION USING EDGE INFORMATION AND TEXTURE INFORMATION
Kyeongseok Jang, Sung Jae Ha and Kwang Chul Son
|
Abstract: In this paper, we propose a deep learning-based SOD model that uses edge and texture information and attention to improve feature loss that occurs when a deep network is configured in Salient Object Detection (SOD). Most of the existing salient object detection models have an autoencoder-based structure, and feature loss occurs in the encoder that extracts and compresses features and the decoder that expands and restores the compressed features. Characteristic loss refers to a result of detecting a background other than an object in an image or not detecting an object. The proposed method uses edge information and texture information to supplement edge information and spatial information of objects that are liable to be lost during learning. Also, feature loss is minimized by emphasizing information that may be lost during learning in a deeply constructed network through spatial attention. As a result of training through the proposed model, mean absolute error (MAE) in the ECSSD dataset was improved by 0.05643 and 0.02547 compared to FCN and U-Net, respectively. The F-measure was improved by 0.04555 and 0.01783, and the MAE was improved by 0.04679 and 0.03503 in the MSRA10K dataset, and 0.02673 and 0.0136 in the F-measure, respectively. |
Keywords and phrases: salient object detection, deep learning, edge information.
|
|
Number of Downloads: 273 | Number of Views: 448 |
|