A One-stage Detector for Extremely-small Objects Based on Feature Pyramid Network

Sammanfattning: Thanks to the recent development in Graphics Processing Unit (GPU) and deep neural network, outstanding enhancement has been made in real-time and multi-scale object detection. However, most of these detectors ignore the situations where the target needs to be identified is extremely-small corresponding to the size of the image or video. The spatial resolution of feature maps is decreasing and detailed information about extremely-small objects is missing during the process of extracting features with stride and pooling. So how to keep the higher spatial resolution when we extract the richer semantic information and enlarge receptive field becomes the crucial core of this project. With the purpose of detecting targets with 30 to 1000 pixels in 1080p videos, we design a one-stage detector that uses DetNet as the backbone and construct the head of detector based on the idea of Feature Pyramid Network (FPN). Taking advantage of the dilated convolutional layer in DetNet, the size of the last three feature maps are not decreasing. By contrast, the receptive field and semantic information are augmented by traversing the backbone of the detector. Besides, with the technique of FPN, feature maps from different stages are combined and assigned to the prediction, making the model more robust and accurate. Additionally, in order to reduce the input size of the image to decrease computational complexity without missing any information of extremely-small objects, we crop the full image based on the distribution of the target’s location in existing data instead of directly resizing the full image. We compare the performance of this proposed detector with YOLOv3 on the custom dataset, and it turns out to obtain remarkably good results on extremely small objects, improving mean average precision by 18%.