Authors: Huijie Zhang*, San Diego State University
Topics: Remote Sensing
Keywords: small object detection, convolutional neural network, deep learning, urban element detection
Session Type: Virtual Paper
Start / End Time: 9:35 AM / 10:50 AM
Room: Virtual 19
Presentation File: No File Uploaded
Recently, deep learning technologies have brought brilliant improvement for object detection in in the field of remote sensing. However, small object detection such as traffic sign, license plate, and other urban furniture is still a challenging task in remote sensing imagery. The study proposes an adaptive convolutional neural network (A-CNN) model to boost small urban element detection performance in mobile mapping system images. First, we design a new architecture by removing the downsampling operation in the third stage of ResNet backbone, to exploit more local information from the image and extract powerful features for small object detection. Second, a novel training sample strategy is utilized to select positive and negative sample adaptively based on the statistical characteristics of generated anchors. In addition, generalized Intersection over Union (GIoU) is incorporated for loss calculation to the object detection framework. We conducted extensive experiments on the published urban element detection (UED) data set. The experimental results demonstrate that our proposed A-CNN model improves detection performance and achieves better results compared with the state-of-the-art methods.