A real-time small target detection network

Target detection based on deep convolutional neural network has achieved excellent performance. However, small target detection is still one of the challenges in the field of computer vision. In this paper, we present an efficient network for real-time small target detection. The proposed network performs feature extraction using a modified Darknet53, while utilizing scale matching strategy to select suitable scales and anchor size for small target detection. In the network, we design an adaptive receptive field fusion module to increase the context information around the small targets by merging the features with different receptive field. Furthermore, we also propose an image cropping method in data preprocessing, aiming to make the targets trained in a wider range of scales. We conduct experiments on VEDAI dataset and small target dataset. Comparative results show that the proposed network achieved 74.5% mean average precision (mAP) at 50.0 FPS on VEDAI dataset and 45.7% mAP at 51.1 FPS on small target dataset which is better than other advanced target detectors.