A new approach based on deep learning for object detection from camouflaged images

The term "camouflage object detection" (COD) is often used to describe the detection of hidden objects in an environment. Since the objects and their environments are very similar and the lack of large datasets in the detection of camouflaged objects makes it difficult to carry out related studies. In this thesis, a new COD approach based on deep learning is proposed. In the first stage, features are extracted from a backbone network (ResNet). Those features are then fed into a proposed inception module, which extracts rich context features from a large receptive field and improves integration between feature maps. It is observed that the proposed module significantly improves the segmentation. In the second step, low-level features are densely combined with features rich in semantic information. In the third step, an attention mechanism is used to better distinguish the camouflaged object from the background. Transform attention, channel attention and spatial attention modules are used to extract more information about the target region and suppress redundant information. Experiments on the modules show that all combinations accurately detect camouflaged objects. Furthermore, the proposed method is tested on widely used datasets and its performance is evaluated. In addition, the proposed model is also evaluated for medical image segmentation and uncamouflaged object detection with acceptable results are obtained.