Lightweight Pixel Difference Networks for Efficient Visual Representation Learning.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Developing lightweight Deep Neural Networks (DNNs) for edge devices requires balancing accuracy and efficiency.
• Existing methods face challenges in optimizing both high accuracy and high computational efficiency.

Purpose of the Study:

• To introduce novel convolutional methods, Pixel Difference Convolution (PDC) and Binary PDC (Bi-PDC), for creating efficient DNNs.
• To present two lightweight networks, PiDiNet and Bi-PiDiNet, that leverage PDC and Bi-PDC for enhanced visual task performance.

Main Methods:

• Proposed Pixel Difference Convolution (PDC) and Binary PDC (Bi-PDC) to capture higher-order local differential information efficiently.
• Developed PiDiNet and Bi-PiDiNet, integrating PDC and Bi-PDC for edge detection and object recognition.
• Conducted extensive experiments on datasets like BSDS500, ImageNet, LFW, and YTF.

Main Results:

• PiDiNet and Bi-PiDiNet demonstrated the best accuracy-efficiency trade-off among tested models.
• PiDiNet achieved human-level edge detection performance on BSDS500 (100 FPS, 1M parameters) without ImageNet pre-training.
• Bi-PiDiNet achieved state-of-the-art accuracy among Binary DNNs for object recognition with reduced computational cost.

Conclusions:

• PDC and Bi-PDC offer a computationally efficient approach to designing accurate and lightweight DNNs.
• PiDiNet and Bi-PiDiNet represent significant advancements in efficient visual representation learning for edge computing.
• The proposed methods enable high-performance visual tasks on resource-constrained devices.