| Summary: | 碩士 === 國立中山大學 === 資訊工程學系研究所 === 107 === Deep learning (DL) is a popular research topic in artificial intelligence (AI) and has been successfully applie to many AI applications including image classification, speech recognition, object detection. DL-based object detection architectures can be divided into two categories: two-stage detector and one-stage detector, depending one whether the two major networks, feature extraction network and region proposal network, are independent or related. Both types of object detection architectures generate image features using DL network, followed by operations of region proposal network. Current research focus of DL-based object detection is on the design of efficient region proposal network, with feature extraction network usually realized by benchmark deep convolutional neural networks (CNN) such as VGG and AlexNet, resulting in slow processing speed due to many layers of CNN operations. This thesis proposal presents the designs of low-complexity CNN feature extraction network with significant improvement of processing speed while maintaining similar accuracy level of object detection. The approaches is applied to one-stage detector of SSD (single shot detector) in DL-based object detection and can achieve real-time processing speed on GPU server without too much loss of detection accuracy.
|
|---|
