光学学报, 2020, 40 (1): 0111018, 网络出版: 2020-01-06
基于轻量级残差网络的红外遥感船只检测 下载: 1645次
Infrared-Remote-Sensing Ship Detection Based on Lightweight Residual Network
成像系统 红外船只检测 二值网络 编码-解码结构 三元残差网络 imaging systems infrared ship detection binary network encoder-decoder structure ternary residual network
摘要
针对红外遥感船只检测领域存在的硬件存储资源和功耗的限制,以及目标检测输出边界矩形框形式结果不够精细的问题,提出了一种轻量化且具有像素级输出的分割网络TRS-Net。将图像分割的编码-解码结构用于船只检测,以获得像素级的输出;将32 bit的浮点型参数二值化(目的是压缩网络模型的大小),提出了BS-Net;针对BS-Net带来的检测精度低的问题,引入残差连接,提出了BRS-Net;根据神经网络稀疏性特点引入参数三元化,提出了TS-Net;为进一步提升检测效果,将TS-Net改进成TRS-Net。采用实验室自主研制的长波红外相机进行成像实验,获取红外船只图片并制作数据集,对4种网络的结果进行对比分析。结果表明:TRS-Net检测的精确率为88.73%,召回率为83.34%,F1-score为85.95%,交并比为75.36%,模型大小压缩为原先的1/16。TRS-Net对红外船只的实时检测具有一定的工程应用价值。
Abstract
To address the limitations of hardware storage resource and power consumption in infrared-remote-sensing ship detection and the inadequate precision of the output boundary rectangular box form of target detection, a lightweight and pixel-level-output segmentation network TRS-Net (ternary residual segmentation network) is proposed. We apply the encoder-decoder structure of image segmentation to ship detection to obtain the pixel-level output. Further, we binarize the 32-bit floating-point parameters to compress the size of the network model and propose a binary segmentation network (BS-net). Then, to solve the problem of poor detection accuracy caused by BS-Net, we introduce residual connection and propose a binary residual segmentation network (BRS-Net). Furthermore, owing to the sparsity of the neural network, we introduce ternary parameters and propose a ternary segmentation network (TS-Net); therefore, we propose a ternary residual segmentation network (TRS-Net) to further improve the detection effect. Using a long-wave infrared camera independently developed by the laboratory for imaging experiments, we obtain infrared images of ships, make the datasets, and compare and analyze the results of four kinds of networks. The results demonstrate that the detection precision, recall rate, F1-score, and intersection-over-union of TRS-Net are 88.73%, 83.34%, 85.95%, and 75.36%, respectively. Furthermore, the model size is reduced to one-sixteenth of its original size. Therefore, the proposed TRS-Net has practical engineering value for real-time infrared ship detection.
朱天佑, 黄凌锋, 董峰, 龚惠兴. 基于轻量级残差网络的红外遥感船只检测[J]. 光学学报, 2020, 40(1): 0111018. Tianyou Zhu, Lingfeng Huang, Feng Dong, Huixing Gong. Infrared-Remote-Sensing Ship Detection Based on Lightweight Residual Network[J]. Acta Optica Sinica, 2020, 40(1): 0111018.