一种基于U-Net语义分割网络的多光谱迷彩目标识别方法

A Multispectral Camouflage Target Recognition Method Based on U-Net Semantic Segmentation Network

提出了一种基于U-Net的多光谱迷彩目标识别方法。设计数据采集方案采集迷彩目标多光谱数据;采用不同尺度卷积核提取联合的光谱空间特征;编码结构中采用残差学习加深网络深度,使网络能学习到更加丰富抽象的特征;对深层的特征图进行上采样与浅层特征图相加增强浅层特征图中的语义信息。与3通道U-Net语义分割网络相比,召回率提高了62.65%,F1-Score提高了50.18%,证明了采用多光谱识别迷彩目标的显著优势;与6通道U-Net语义分割网络相比,精确率保持基本不变的同时召回率提高了3.42%,F1-Score提高了1.62%,在保证检测准确的前提下进一步减少了误检。

In order to improve the ability of camouflage target recognition,a multi-spectral camouflage target recognition method based on U-Net is proposed. A data collection scheme is designed to collect multi-spectral data sets of camouflage targets.Different scale convolution kernels are used to extract joint spectral-spatial features. Residual learning in the coding structure is used to deepen the network depth,so that the network can learn more rich and abstract features. Upsampling-deep feature is added to shallow feature maps to enhance semantic information in shallow feature maps. Compared with the 3-channel U-Net,the Recall is increased by 62.65%,and the F1-Score is increased by 50.18%,demonstrating the significant advantages of using multi-spectral recognition camouflage targets. Compared with the 6-channel U-Net,the accuracy rate remains basically the same while the recall increases by 3.42%,and the F1-Score increases by 1.62%,further reducing false detections on the premise of ensuring precision.