Recognition and Backbones

This database tracks the backbone lineage of computer vision: how visual representations moved from early CNNs to very deep residual networks, efficient mobile architectures, neural architecture search, and modernized ConvNets.

Early CNNs and ImageNet Era

Year	Paper	Topic	Note
1998	Gradient-Based Learning Applied to Document Recognition	LeNet	Early end-to-end CNN for document and digit recognition.
2012	ImageNet Classification with Deep Convolutional Neural Networks	AlexNet	Large-scale CNN breakthrough on ImageNet.
2013	Visualizing and Understanding Convolutional Networks	ZFNet	Interprets CNN features and refines AlexNet-style design.
2014	Very Deep Convolutional Networks for Large-Scale Image Recognition	VGG	Deep stack of small 3x3 convolutions.
2014	Going Deeper with Convolutions	GoogLeNet / Inception	Multi-branch convolutional modules for efficient depth.

Depth, Connectivity, and Scaling

Year	Paper	Topic	Note
2015	Batch Normalization	Training stability	Normalization layer that helped deep CNNs train reliably.
2015	Deep Residual Learning for Image Recognition	ResNet	Residual connections made very deep networks practical.
2016	SqueezeNet	Small CNNs	AlexNet-level accuracy with far fewer parameters.
2016	Densely Connected Convolutional Networks	DenseNet	Dense feature reuse and improved gradient flow.
2016	Aggregated Residual Transformations for Deep Neural Networks	ResNeXt	Adds cardinality as a capacity axis.
2017	MobileNets	Efficient CNNs	Depthwise separable convolutions for mobile vision.
2019	EfficientNet	Compound scaling	Balances depth, width, and resolution.
2020	Designing Network Design Spaces	RegNet	Systematic design space for efficient ConvNets.
2022	A ConvNet for the 2020s	ConvNeXt	Modernized ConvNet competing with Vision Transformers.

Reading Path

Step	Read
1	LeNet and AlexNet for the pre-ImageNet to ImageNet transition.
2	VGG and Inception for depth and module design.
3	BatchNorm and ResNet for trainability.
4	DenseNet and ResNeXt for connectivity and capacity.
5	MobileNet, EfficientNet, RegNet, and ConvNeXt for efficiency and modern backbone design.