摘要
在视觉物体识别任务中，光照变化会导致物体外观发生显著改变，从而干扰基于深度神经网络的识别模型。尤其在某些罕见光照条件下，获取足够多的训练样本往往耗时且成本高昂。为解决这一问题，本文提出一种新型神经网络架构——分离光照网络（Separating-Illumination Network, Sill-Net）。Sill-Net能够学习将图像中的光照特征进行分离，并在训练过程中，将这些分离出的光照特征在特征空间中用于增强训练样本。实验结果表明，所提方法在多个物体分类基准测试中均优于当前最先进的技术。
代码仓库
lanfenghuanyu/Sill-Net
官方
pytorch
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| few-shot-image-classification-on-cifar-fs-5 | Illumination Augmentation | Accuracy: 87.73 |
| few-shot-image-classification-on-cifar-fs-5-1 | Illumination Augmentation | Accuracy: 91.09 |
| few-shot-image-classification-on-cub-200-5 | Illumination Augmentation | Accuracy: 96.28 |
| few-shot-image-classification-on-cub-200-5-1 | Illumination Augmentation | Accuracy: 94.73 |
| few-shot-image-classification-on-mini-2 | Illumination Augmentation | Accuracy: 82.99 |
| few-shot-image-classification-on-mini-3 | Illumination Augmentation | Accuracy: 89.14 |
| traffic-sign-recognition-on-belgalogos | Sill-Net | Accuracy: 89.48 |
| traffic-sign-recognition-on-belgian-traffic | Sill-Net | Accuracy: 98.97 |
| traffic-sign-recognition-on-chinese-traffic | Sill-Net | Accuracy: 97.19 |
| traffic-sign-recognition-on-flickrlogos-32 | Sill-Net | Accuracy: 95.80 |
| traffic-sign-recognition-on-gtsrb | Sill-Net | Accuracy: 99.68% |
| traffic-sign-recognition-on-toplogo-10 | Sill-Net | Accuracy: 89.66 |
| traffic-sign-recognition-on-tsinghua-tencent | Sill-Net | Accuracy: 99.53 |