Res2Net

Res2Net provides enhanced multi-scale feature representation in convolutional neural networks by constructing hierarchical residual-like connections within a single residual block to capture features at multiple scales with expanded receptive fields.

Key Features:

• Hierarchical Multi-Scale Representation: Constructs multi-scale feature representations inside individual residual blocks via hierarchical residual-like connections.
• Expanded Receptive Fields: Increases the range of receptive fields for each network layer to capture larger contextual information.
• Modular Integration: The Res2Net block can be integrated into backbone CNN models such as ResNet, ResNeXt, and DLA.

Scientific Applications:

• Image Classification: Demonstrated performance improvements on CIFAR-100 and ImageNet compared with baseline models.
• Object Detection: Applied to object detection tasks with reported performance gains.
• Class Activation Mapping: Shown advantages in class activation mapping for localizing discriminative regions.
• Salient Object Detection: Applied to salient object detection with confirmed benefits.

Methodology:

Constructs hierarchical residual-like connections within a single residual block to produce granular multi-scale feature representations and expands per-layer receptive fields; the Res2Net block is integrated into backbone CNNs such as ResNet, ResNeXt, and DLA.