Efficient Light-Weight Deep Neural Network for Person Detection in Drone Images

Abstract

In this paper, we propose an efficient light-weight deep neural network model for small object (person) detection in drone images. The proposed method performs light-weight as well as efficient small object detection by removing the head layers that detects large and medium-sized objects. In addition, the feature was extracted by focusing the weight on the small object while performing feature fusion through the Weighting Module. Finally, since the class imbalance problem between the object and the background is more serious in the drone image, the problem is alleviated by using the focal loss. As a result, the light-weight that can be mounted on the drone and the inference time are faster, and the Average Precision (AP) is higher than the original model.