In this work we investigate the effect of the convolutional network depth on its accuracy in the large-scale image recognition setting. Our main contribution is a thorough evaluation of networks of increasing depth using an architecture with very small (3x3) convolution filters, which shows that a significant improvement on the prior-art configurations can be achieved by pushing the depth to 16-19 weight layers. These findings were the basis of our ImageNet Challenge 2014 submission, where our team secured the first and the second places in the localisation and classification tracks respectively. We also show that our representations generalise well to other datasets, where they achieve state-of-the-art results. We have made our two best-performing ConvNet models publicly available to facilitate further research on the use of deep visual representations in computer vision.
核心贡献 · Key contributions
系统评估了使用极小 3x3 滤波器的卷积网络深度增加的效果。 Systematic evaluation of increasing depth in ConvNets using very small 3x3 filters.
证明将深度推至 16-19 个权重层可显著提升准确率。 Demonstrated that pushing depth to 16-19 weight layers significantly improves accuracy.
在 ILSVRC 2014 定位和分类赛道分别获得第一和第二名。 Achieved first and second places in ILSVRC 2014 localization and classification tracks.
表明学到的表示能很好地泛化到其他数据集，取得最优结果。 Showed that learned representations generalize well to other datasets, achieving state-of-the-art results.
公开了两个性能最佳的卷积网络模型以促进进一步研究。 Publicly released two best-performing ConvNet models to facilitate further research.
确认了深度对视觉表示质量的关键作用。 Confirmed that depth is crucial for visual representation quality.
局限 · Limitations
评估仅限于 ImageNet 数据集，未充分探索对其他领域的泛化。 Evaluation limited to ImageNet dataset; generalization to other domains not fully explored.
超过 19 层的深度未提升性能，观察到饱和现象。 Depth beyond 19 layers did not improve performance; saturation observed.
训练需要大量算力（4 块 GPU 上 2-3 周）。 Training required significant computational resources (2-3 weeks on 4 GPUs).
架构仅使用 3x3 滤波器，未系统比较其他滤波器尺寸。 Architecture uses only 3x3 filters; other filter sizes not systematically compared.
多裁剪评估增加计算时间，但准确率提升不成比例。 Multi-crop evaluation increases computation time without proportional accuracy gain.
论文章节 · Sections（共 17）
摘要Abstract
1 引言1 Introduction
2 卷积网络配置2 ConvNet Configurations
2.1 架构2.1 Architecture
2.2 配置2.2 Configurations
2.3 讨论2.3 Discussion
3 分类框架3 Classification Framework
3.1 训练3.1 Training
3.2 测试3.2 Testing
3.3 实现细节3.3 Implementation Details
4 分类实验4 Classification Experiments
4.1 单尺度评估4.1 Single Scale Evaluation
4.2 多尺度评估4.2 Multi-Scale Evaluation
4.3 多裁剪评估4.3 Multi-crop evaluation
4.4 卷积网络融合4.4 ConvNet Fusion
4.5 与最新技术比较4.5 Comparison with the State of the Art