State-of-art object detection networks like YOLO, SSD and Faster R-CNN all have achieved great success in object detection. However, these algorithms have a low performance in small object detection. So, we produce the Expanding receptive field YOLO (ERF-YOLO) to deal with this problem. At first, we propose an efficient block which is called expanding receptive field block (ERF-block) to capture more information in larger areas. Base on YOLOv2, we down-sample the low-level location information by ERF-block, and up-sample feature information by deconvolution. Then we further assemble these two parts together to make the prediction. After training the network on VOC dataset, we have a good result with 82.6% mAP (mean Average Precision) which is 4.0% higher than the original YOLOv2 network. Thanks to the efficient block, it takes 62fps to detect one image when the input size is 416×416, which could keep a real-time speed. In addition, we also evaluate the model on a remote sensing dataset which contains many small targets, and it also shows that ours model has a better performance.

No abstract

Partial reconfigurable system is an architecture consisting general purpose processors and FPGAs, in which FPGA can be reconfigured in run-time. Based on the architecture, software tasks and hardware tasks that are executed on processor and FPGA respectively co-exist. In this paper, a real-time fault-tolerant scheduling algorithm is proposed to schedule software/hardware hybrid tasks. In the algorithm, the sufficient condition for schedulable hybrid tasks is derived from analyzing system operation conditions when the first deadline is missed, and rollback/recovery and TMR approaches are used respectively to schedule software subtasks and hardware subtasks for fault tolerance. The experimental results demonstrate that all deadlines of accepted hybrid tasks are met and processor's utilization ratio is increased greatly compared with that of the exiting approaches when multiple faults occur.

No abstract

A deep learning model has a large number of free parameters, which need to be effectively trained on a large number of samples to calculate the depth parameters. However, many special applications like underwater acoustic signal recongnition cannot provide enough dataset to guarantee high performance. In addition, the original dataset adopts some formats, such as audio, which makes it difficult to capture features. To overcome these challenges, we propose a novel framework. Firstly, based on the evaluation of spectrum method, our framework selects the appropriate preprocessing method. Then it modifies GAN to generate samples, and establishes an independent classification network to ensure the quality of samples. Finally, our framework applies the existing classification network to evaluate performance and selects the best one for the LOFAR spectrum. The experimental results demonstrate that the proposed method can generate high-quality LOFAR spectrum and improve the prediction accuracy of the classification model significantly.