Anabranch network for camouflaged object segmentation

D$^{2}$C-Net: A Dual-Branch, Dual-Guidance and Cross-Refine Network for Camouflaged Object Detection

“…In recent years, to accurately distinguish the camouflaged objects in complex backgrounds, increasing architectures apply convolutional neural networks (CNNs) to camouflaged object detection. For example, Nie et al [21] proposed an endto-end network (ANet) based on CNNs to precisely separate the camouflaged objects from given images. In [20], Zheng et al presented a dense deconvolution network (DDCN) for camouflaged object detection to effectively detect the target object in the background of artificial camouflage.…”

“…CAMO [21] was proposed in 2019 as the first dataset for camouflaged object segmentation, which consists of 1250 images and is divided into two categories (artificially camouflaged objects and naturally camouflaged objects). 1000 images are used for training and the rest for testing.…”

“…For example, to explore the detection of human who wears camouflage pattern in complex background scenes, Zheng et al [20] employed a bit of short connections into the framework that is beneficial to predict accurate camouflage maps. Le et al [21] proposed a two-stream (classification stream and detection stream) framework for camouflage object detection. Recently, Fan et al [22] proposed SINet, which simulates the human visual mechanism.…”

“…Considering the different contributions of low-level features and high-level features to generate the camouflage maps [21]- [23], and due to the high-resolution of low-level features, we employ the third layer of the backbone as the node to simulate the two stages of the human eye observing the camouflage scene. The first stage is simulated to the human observation of potential camouflage objects in the scene.…”

“…3) Benefiting from the peer features can provide some cues about the location of target object regions, we employ a self-refine attention unit and cross-refine unit to conduct more accurate camouflage maps. 4) To prove the superiority of our model, we compare our model with other 13 state-of-the-art models on three benchmark datasets (CAMO [21], CHAMELEON [24], COD10K [5]). Our model acquires the best performance basing on all five public evaluation metrics.…”