Real-Time Target Detection in Visual Sensing Environments Using Deep Transfer Learning and Improved Anchor Box Generation

Visual perception is critical and essential to understand phenomenon and environments of the world. Pervasively configured devices like cameras are key in dynamic status monitoring, object detection and recognition. As such, visual sensor environments using one single or multiple cameras must deal with a huge amount of high-resolution images, videos or other multimedia. In this paper, to promote smart advancement and fast detection of visual environments, we propose a deep transfer learning strategy for real-time target detection for situations where acquiring large-scale data is complicated and challenging. By employing the concept of transfer learning and pre-training the network with established datasets, apart from the outstanding performance in target localization and recognition can be achieved, time consumption of training a deep model is also significantly reduced. Besides, the original clustering method, k-means, in the You Only Look Once (YOLOv3) detection model is sensitive to the initial cluster centers when estimating the initial width and height of the predicted bounding boxes, thereby processing large-scale data is extremely time-consuming. To handle such problems, an improved clustering method, mini batch k-means++ is incorporated into the detection model to improve the clustering accuracy. We examine the sustainable outperformance in three typical applications, digital pathology, smart agriculture and remote sensing, in vision-based sensing environments.