Wildlife Small Object Detection based on Enhanced Network in Ecological Surveillance

Visual tele-observation is an effective way for intelligent monitoring and protection of ecology and the natural environment. Different from pedestrian or rigid body detection, wildlife detection in natural scenes face more complex problems, such as the existence of wild background clutter, local or global vegetation occlusion of animals, small object, rotation, deformation, and other interfering factors. In this paper, We mainly propose a novel method for the small object problem. For our small object wild species data set, we use an SSD detector for object detection. Firstly, the K-means algorithm is used to adjust the anchor box of the SSD network. Secondly, for the situation that SSD is not good for small object detection, in conv4_3 layers of SSD, the feature enhancement module is added. Finally, aiming at the size of a small object, the network level of SSD is deleted. We have proved the feasibility of each of the three methods through experiments and combined with the three methods to verify, the recognition rate of the small target has increased by 2.67%.