ResNet50

ResNet50 is a 50-layer deep convolutional neural network in the ResNet family, introduced by Kaiming He and colleagues in 2015. It uses residual learning through skip connections to address optimization difficulties in very deep networks, enabling successful training of substantially deeper models. ResNet50 achieved top performance on ImageNet and has become a widely adopted backbone for visual recognition tasks.

Architecture: The model begins with a 7x7 convolution with stride 2 and a max-pooling layer, followed by four stages of bottleneck blocks labeled conv2_x through conv5_x. In ResNet50, the stages contain 3, 4, 6, and 3 blocks respectively. Each bottleneck block uses a 1x1 reduction, a 3x3 core, and a 1x1 expansion, producing 256, 512, 1024, and 2048 channels thereafter. The network ends with global average pooling and a final fully connected layer for 1000-class ImageNet classification.

Key features: Residual connections add the input of a block to its output, enabling identity mappings when dimensions match and facilitating gradient flow during training. The architecture employs batch normalization and ReLU activations, and is typically trained with stochastic gradient descent and standard data augmentation.

Variants and usage: ResNet50 serves as a general-purpose backbone for image classification and serves as a feature extractor for object detection and segmentation. Deeper variants such as ResNet101 and ResNet152 extend the same bottleneck design, and pre-trained weights for ResNet50 are widely available in major frameworks.

See also: ResNet family.