搜索

x
中国物理学会期刊

基于VCSEL-SA光子脉冲神经元灰度形态学运算与白顶帽变换的红外小目标增强

Infrared Small-Target Enhancement Based on Photonic Gray-Scale Morphological Operation and White Top-Hat Transform in a VCSEL-SA Spiking Neuron Network

PDF
导出引用
  • 本文结合带饱和吸收体的垂直腔面发射激光器(vertical-cavity surface-emitting laser with a saturable absorber,VCSEL-SA)的优势,提出并数值研究了一种基于灰度形态学运算和白顶帽变换实现红外小目标增强的光子神经网络。研究结果表明:将灰度图像阈值分解后的二值图像和结构元素编码合并后注入到VCSEL-SA可实现形态学开运算。将原始图像减去由各阈值层图像形态学开运算结果逐层叠加重构的背景估计图像即可完成白顶帽变换,实现小目标增强。通过对VCSEL-SA尖峰动力学的分析确定了可行的脉冲编码条件。在信噪比SNR=10 dB和时间抖动σt=80 ps时,VCSEL-SA对编码脉冲的尖峰响应准确率仍可保持在97%以上。对于灰度米粒图像,处理后米粒图像的信杂比增益(SCRG)、信噪比增益(SNRG)、背景抑制因子(BSF)可分别达到1.77,1.46和2.04。在NUDT-SIRST数据集测试中,基于不同结构元素的小目标灰度图像增强后的平均SCRG、SNRG和BSF最大值可分别达到7.73、2.67和16.34。通过分析不同目标尺度及SCR下最优结构元素尺度的统计规律,并结合原始图像的局部灰度统计特征可以初步确定最优的结构元素尺度,获得优于传统方法的小目标增强效果。研究结果可为基于VCSEL-SA光子神经网络的红外小目标增强和背景抑制提供一种可行的光子实现方法。

     

    Infrared small-target enhancement is of great significance in fields including infrared search, remote sensing, surveillance, and early warning. In single-frame infrared images, small targets usually occupy few pixels and are characterized by weak intensity, a low signal-to-clutter ratio, and limited texture information, rendering them susceptible to background clutter, nonuniform background fluctuations, strong edges and random noise. Gray-scale morphological operation and the white top-hat transform provide an effective means for background suppression and local bright small-target enhancement. However, conventional electronic implementations involve repeated sliding-window operations, neighborhood access, local comparison and intermediate data access. In this work, we propose and numerically investigate a photonic spiking neural network based on a vertical-cavity surface-emitting laser with a saturable absorber (VCSEL-SA), which can implement gray-scale morphological operation and white top-hat transform for infrared small-target enhancement. Firstly, gray-scale image is decomposed into a set of binary threshold images. Secondly, for each threshold image, the local image window and the structuring element are unfolded according to the same spatial order and temporally encoded into two rectangular optical-pulse sequences. Then, the two encoded pulse sequences are combined with a certain delay, and injected into the VCSEL-SA to implement the morphological operation including erosion and dilation according the fired spiking number. Subsequently, the binary opening operation results are stacked to reconstruct the new gray-scale image, which is used as the background estimate. As a result, small-target enhancement can be achieved by subtracting this background estimate from the original image. The spiking dynamics of VCSEL-SA is investigated to determine the feasible pulse-encoding conditions. Under Gaussian white noise with a signal-to-noise ratio (SNR) of 10 dB and input-pulse timing jitter with a standard deviation of 80 ps, the spike-response accuracy of VCSEL-SA to injected encoding pulses remains above 97%. For a gray-scale rice image, the signal-to-clutter ratio gain (SCRG), signal-to-noise ratio gain (SNRG), and background suppression factor (BSF) after small-target enhancement processing can reach 1.77, 1.46, and 2.04, respectively. The proposed method is further evaluated on the NUDT-SIRST dataset using structuring elements with different shapes and sizes. For the tested structuring elements, the maximal SCRG, SNRG, and BSF values can reach 7.73, 2.67 and 16.34, respectively. After analyzing the statistic regularities of optimized structuring-element size under different target sizes and SCRs, the optimized structuring-element size required by the small-target enhancement can be preliminarily determined based on the local statistical characteristics of original gray-scale image. Compared with traditional small-target image-enhancement methods, this proposed method can achieve relatively good indices. These results provide a feasible photonic implementation method for infrared small-target enhancement and background suppression based on a VCSEL-SA photonic spiking neural network.

     

    目录

    /

    返回文章
    返回