Small Object Detection From Drone Perspective

Object detection is an important part of computer vision.In object detection,small object detection is a challenging task because scale of object is small and resolution of object is low.The demand for small object detection in images from drone perspective is growing.In this thesis,we propose several novel methods for small object detection.The main works are summarized as follows:(1)We propose a novel object detection method where the weights used in the loss function are adopted based on the scale of object,and we construct a hybrid data set by combining existing public data sets and self-labeling data sets because of shortcomings of existing public data sets.In this paper,we analyze the relationship between scale size and scale loss.We also explore the relationship between scale size and scale loss in mathematics in a given situation.Experiments in this thesis show that our method has a better performance,and our method improves 2% in Average Precision(AP).(2)We propose a cascade object detection to solve the imbalance between positive sample and negative sample.Because boosting algorithm can improve small object detection,we use boosting algorithm in small object detection by adjusting the loss function in training.The distribution of positive sample and negative sample is unbalanced in small object detection.So we change the loss of negative sample to solve the problem.We modify the structure of network,because the network is not suitable for small object detection.In experiments,the results shows that our method performs better by using the three methods mentioned above.Our method improves 1.34% in AP.(3)We propose a method in data augmentation for small object detection.The number of small objects in data set is very small.We solve the problem by data augmentation.We introduce some traditional methods and advanced methods in data augmentation.Inspired by the advanced methods.we generate new data by object fusion in image.In experiments,we add the new data to train set.It improves 3% in AP compared with original train set.