A novel residual block: replace Conv1× 1 with Conv3×3 and stack more convolutions.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Residual structures are crucial for neural network performance in computer vision.
• Classical residual networks face limitations in receptive field size, channel, and spatial information processing.

Purpose of the Study:

• To propose a novel residual structure that overcomes limitations of classical residual networks.
• To enhance effective receptive fields, channel information fusion, and spatial feature extraction.

Main Methods:

• Modification of identity mapping and down-sampling blocks within the residual structure.
• Ablation studies to verify performance in channel information fusion and spatial feature extraction.
• Design and testing of a non-deep convolutional neural network (CNN) using the novel structure on Cifar10 and Cifar100 datasets.

Main Results:

• The novel residual structure demonstrates superior performance in channel information fusion and spatial feature extraction.
• The proposed network model achieved higher accuracy than mainstream networks (ResNet50, ResNeXt50, SeResNet152) on benchmark datasets.
• Achieved 3.08% higher accuracy than ResNet50 and 1.38% higher than ResNeXt50.
• Outperformed SeResNet152 by 0.29% with 50 fewer training epochs.

Conclusions:

• The proposed novel residual structure significantly improves neural network performance for computer vision tasks.
• The enhanced effective receptive field and feature extraction capabilities offer a competitive alternative to existing architectures.
• The model shows promise for efficient and high-accuracy image recognition applications.