Object Segmentation based on Anti-disturb Network in Ecological Surveillance

Visual tele-observation is an effective way for ecological protection. Different from pedestrian or rigid body detection, wildlife recognition in natural scenes face more complex problems, such as the similarity of background and foreground, wild background clutter, deformation, and other factors, which will lead to a reduction in recognition accuracy. In this paper, We propose a novel segmentation network that is named T-S network to segment background and foreground. The network mainly consists of two parts: one is a tracking network. Considering the context similarity of images in the data set, the purpose of tracking is to ensure that the position of the object is accurately known when the same species in the images are synthesized into a video with interval frames; the other is the saliency network, which is the basis of the most likely position of the object obtained from the tracking network. Then, the saliency network is used to complete the object segmentation. To verify the effectiveness of the network and consider the size of the data set, a lightweight CNN is used to identify the network. Experiments show that the accuracy of task recognition is improved by 1.89% by using the segmentation method proposed in this paper.