TY - JOUR
T1 - CENet: Content-Aware and Edge-Aware Network for Salient Object Detection
AU - Gao , Zhikuan
JO - Journal of Information and Computing Science
VL - 1
SP - 028
EP - 039
PY - 2024
DA - 2024/01
SN - 17
DO - http://doi.org/
UR - https://global-sci.org/intro/article_detail/jics/22359.html
KW - Saliency object detection, Content-aware and edge-aware feature, Axial-attention transformer, Learning-based feature fusion one.
AB - <p style="text-align: justify;">Benefitting from Fully Convolutional Networks (FCNs), salient object detection methods have
achieved prominent performance. However, there are still some challenges in this task: 1) lack of effective
feature representation and integration make the result salient maps lose some regions of object, or bring some
non-saliency regions. 2) suffering from the continuous pooling or stride operations, the predicted maps will
lose some important spatial detail information, especially the boundary of object. To address these two
problems, we propose the Content-aware and Edge-aware network (CENet) which contains three sub-modules:
1) we design a content-aware feature extraction module which uses a transformer block and channel-wise
attention mechanism to capture the distinct content features and suppresses the non-saliency regions. 2) an
edge-aware feature extraction module is introduced to learn the boundary features and predict the intact edge of
the salient object. 3) a feature fusion module is proposed to integrate features from the first two module in a
learning way. We also design a hybrid loss function which has better performance than the widely-used binary
cross entropy loss. Results show that, our method can detect the intact salient object without losing regions of
object or bring some non-saliency regions, and can also obtain the precise boundary. Experimented on several
datasets, our method can achieve the state-of-art performance.</p>