Salient Object Detection based on Deep Autoencoder Network with ELU Residual Block

Abstract

In this paper, we proposed a deep autoencoder segmentation method using ELU residual block and concatenation to reduce the loss of features and improve the accuracy by salient object detection based on deep learning. The existing saliency detection and segmentation methods have an Autoencoder structure, and many features are lost in the process of extracting and compressing features, and the process of expanding and restoring the compressed features. These losses indicate that the background was segmentation, or the object was not segmentation. In the Encoder process, which was a feature extraction stage for improving such a case, detailed information was utilized through skip connection of a residual block, and loss of features is prevented by using an ELU as an activation function. After feature extraction in Encoder process, feature loss occurs because feature was simply expanded in process of Decoder. In order to prevent these losses, the features generated in the process of Encoder were connected to concatenate to utilize in Decoder. The proposed method reduced the loss of features and improved salient object detection in the Autoencoder structure. The proposed method showed improved results compared to the existing method.