FasterMLP efficient vision networks combining attention mechanisms and wavelet downsampling.

Area of Science:

• Computer Vision
• Deep Learning Architectures
• Neural Networks

Background:

• Existing models integrate Multi-layer Perceptrons (MLPs), Convolutional Neural Networks (CNNs), and attention mechanisms for enhanced computer vision performance.
• Resource-constrained and real-time applications require computationally efficient yet accurate models.

Purpose of the Study:

• To propose FasterMLP, a novel lightweight neural network architecture.
• To achieve high computational efficiency and accuracy for computer vision tasks.

Main Methods:

• FasterMLP combines CNN local connectivity with MLP global feature representation.
• Incorporates Convolutional Block Attention Module for enhanced feature extraction.
• Utilizes Haar wavelet downsampling for efficient spatial dimension reduction.

Main Results:

• FasterMLP-S achieves 3.9% higher top-1 accuracy than MobileViT-XXS on ImageNet-1K, with 2x and 2.7x speedup on GPU and CPU.
• FasterMLP-L shows comparable performance to FasterNet-L on COCO with fewer parameters.
• Achieves 81.7% mIoU on Cityscapes, outperforming CCNet and DANet.

Conclusions:

• FasterMLP effectively balances computational efficiency and accuracy.
• Suitable for visual perception in resource-constrained, real-time environments like autonomous driving.