Real-time infrared small target detection network and accelerator design

Abstract

Infrared small target detection (ISTD) plays an important role in military and civil fields. However, various complex backgrounds in real scenarios significantly degrade the ISTD performance. This study proposes an elaborate convolutional neural network (CNN) for ISTD, called small target (ST)-Net, by modeling the task as an image-to-image translation problem. The input image is translated into a clean output map, where only the target remains. ST-Net is a lightweight network with 13 convolutional layers. It uses residual connections to smoothen the forward and backward propagation. The proposed detection method is validated on two different types of widely used infrared small target datasets. Extensive experiments demonstrate that ST-Net could effectively boost the number of correct detections and reduce the number of missed and false detections. Furthermore, a quantized model training method is presented to find as many circuit design benefits as possible while ensuring that the performance degradation of the model is within an acceptable range. Then a high-throughput and energy-efficient accelerator for ST-Net is designed based on the binarized model. Simulation results under CMOS 28 nm technology show that the proposed accelerator can achieve 56 FPS, which meets the real-time requirement, while the power consumption is as low as 48.7 mW. In addition, the accelerator is implemented on FPGA, which can be deployed into an embedded platform.