2018年公认的最佳图像分类神经网络是哪一款？

Great question! Let's walk through this clearly, since the landscape shifted a lot after 2017.

Before 2017, the ImageNet ILSVRC classification challenge was the gold standard for identifying the best CNN image classification models—think AlexNet, VGG, ResNet, all breakout stars from this competition. But as you noted, 2017 split the contest into separate tasks, and by 2018 it moved to Kaggle with a focus on 3D detection, leaving no official top-tier competition for image classification.

That said, in 2018, the field still looked to ImageNet benchmark results to gauge state-of-the-art performance. The most widely recognized top classification models that year were:

• EfficientNet-B0: Developed by Google, this model introduced compound scaling—balancing width, depth, and resolution of the network—to achieve remarkable Top-1 accuracy on ImageNet while being far more computationally efficient than previous large models. It quickly became a go-to baseline for both research and industry thanks to this efficiency-accuracy balance.
• NASNet-A: Another Google project, this model was automatically generated using neural architecture search (NAS). It outperformed many manually designed models on ImageNet, proving that automated architecture design could rival human expertise, and was a big talking point in 2018.
• ResNeXt-101 (32x4d): An extension of ResNet, this model used grouped convolutions to boost expressive power without a massive increase in parameters. It remained a reliable high-performance choice for scenarios where computational resources were less of a constraint.

Among these, EfficientNet stood out as the most impactful of 2018, since its scaling strategy set a new standard for building efficient yet accurate models that's still used today.

内容的提问来源于stack exchange，提问作者SeanJ