مقاله انگلیسی رایگان در مورد عملکرد روش های یادگیری عمیق در تصاویر هوایی – IEEE 2018

Object detection is one of the crucial tasks in computer vision. In the past few years, the performance of object detection [1-14] has dramatically improved due to the success of deep convolutional neural networks (CNN). Typically, object detection and recognition involve two steps: first, deep neural networks are used to localize the potential location of each target object; then, objects are classified into appropriate classes. If the first step can effectively localize the potential object, the second step will be easier. Even though the two-step approach achieved state-of-the-art performance, the running times are usually slow [11]. Therefore, one-stage detectors have been developed to improve the speed. Small-object detection remains challenging because small objects usually have lower resolution and less context information. Finding a 20 × 20 size object located in a 5000 × 5000 image is a difficult task, even for humans. As described in the literature, state-of-the-art methods for object detection usually performed poorly on small objects [11]. Recent research has shown the importance of context information and scale for small-object recognition [8][9]. In addition, it has been reported that lower-layer features extracted from CNNs are very useful for small-object detection and segmentation [8-11]. The work presented in this paper focuses on low-resolution small-object detection by evaluating the performances of leading deep learning methods using a common dataset, which is a new dataset for bird detection, called Little Birds in Aerial Imagery (LBAI). This dataset was created from real-life aerial imagery data, provided by the Illinois Natural History Survey at the University of Illinois at Urbana-Champaign. LBAI contains images of waterfowl and other water birds in shallow lakes within the Illinois River Valley. LBAI includes different colors, shapes, poses, resolutions, and bird sizes range from 10px to 40px. The dataset contains different backgrounds of rivers, vegetation, land, and mixtures between each type of background. Overall, LBAI captures the diversity of real-life situations for bird detection in shallow lake and wet lands across the Midwest. Some of the birds have larger sizes, in higher resolutions and homogenous backgrounds, which make them easier to be identified. While others have smaller sizes, in lower resolutions with blurry contours, making them hard to be detected. LBAI is designed to identify the difficulties and improve existing methods on small object detection. Using the LBAI dataset, we compared a wide-range of representative state-of-the-art deep learning methods. The results shed a light on the strengths and weaknesses of different deep neural network architectures for small object detection. The contributions of this research include applying and adapting leading deep-learning methods to the LBAI dataset, evaluating performances of these methods on a common benchmark dataset for small-object detection and segmentation, and automating the time-consuming process of manual image processing from waterfowl surveys.