Deep learning for small object detection in images

[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT REQUEST OF AUTHOR.] With the rapid development of deep learning in computer vision, especially deep convolutional neural networks (CNNs), significant advances have been made in recent years on object recognition and detection in images. Highly accurate detection results have been achieved for large objects, whereas detection accuracy on small objects remains to be low. This dissertation focuses on investigating deep learning methods for small object detection in images and proposing new methods with improved performance. First, we conducted a comprehensive review of existing deep learning methods for small object detections, in which we summarized and categorized major techniques and models, identified major challenges, and listed some future research directions. Existing techniques were categorized into using contextual information, combining multiple feature maps, creating sufficient positive examples, and balancing foreground and background examples. Methods developed in four related areas, generic object detection, face detection, object detection in aerial imagery, and segmentation, were summarized and compared. In addition, the performances of several leading deep learning methods for small object detection, including YOLOv3, Faster R-CNN, and SSD, were evaluated based on three large benchmark image datasets of small objects. Experimental results showed that Faster R-CNN performed the best, while YOLOv3 was a close second. Furthermore, a new deep learning method, called Retina-context Net, was proposed and outperformed state-of-the art one-stage deep learning models, including SSD, YOLOv3 and RetinaNet, on the COCO and SUN benchmark datasets. Secondly, we created a new dataset for bird detection, called Little Birds in Aerial Imagery (LBAI), from real-life aerial imagery. LBAI contains birds with sizes ranging from 10 by 10 pixels to 40 by 40 pixels. We adapted and applied several state-of-the-art deep learning models to LBAI, including object detection models such as YOLOv2, SSH, and Tiny Face, and instance segmentation models such as U-Net and Mask R-CNN. Our empirical results illustrated the strength and weakness of these methods, showing that SSH performed the best for easy cases, whereas Tiny Face performed the best for hard cases with cluttered backgrounds. Among small instance segmentation methods, U-Net achieved slightly better performance than Mask R-CNN. Thirdly, we proposed a new graph neural network-based object detection algorithm, called GODM, to take the spatial information of candidate objects into consideration in small object detection. Instead of detecting small objects independently as the existing deep learning methods do, GODM treats the candidate bounding boxes generated by existing object detectors as nodes and creates edges based on the spatial or semantic relationship between the candidate bounding boxes. GODM contains four major components: node feature generation, graph generation, node class labelling, and graph convolutional neural network model. Several graph generation methods were proposed. Experimental results on the LBDA dataset show that GODM outperformed existing state-of-the-art object detector Faster R-CNN significantly, up to 12% better in accuracy. Finally, we proposed a new computer vision-based grass analysis using machine learning. To deal with the variation of lighting condition, a two-stage segmentation strategy is proposed for grass coverage computation based on a blackboard background. On a real world dataset we collected from natural environments, the proposed method was robust to varying environments, lighting, and colors. For grass detection and coverage computation, the error rate was just 3%.