Lightweight Pixel Difference Networks for Efficient Visual Representation Learning

TL;DR

This work addresses the trade-off between accuracy and efficiency in visual representation learning for edge devices. It introduces Pixel Difference Convolution (PDC) and Binary Pixel Difference Convolution (Bi-PDC), which inject higher-order local differential information into CNNs and remain differentiable and efficient. It then presents PiDiNet for edge detection and Bi-PiDiNet for object recognition, achieving state-of-the-art or competitive accuracy with far fewer parameters and operations, even without ImageNet pretraining; the authors also provide a frequency-domain interpretation, extensive ablations, and demonstrate generalization to detection tasks with a public code release. Overall, the approach substantially improves the efficiency-accuracy balance for edge-style and recognition tasks and offers a versatile framework for deploying lightweight vision models on constrained devices.

Abstract

Recently, there have been tremendous efforts in developing lightweight Deep Neural Networks (DNNs) with satisfactory accuracy, which can enable the ubiquitous deployment of DNNs in edge devices. The core challenge of developing compact and efficient DNNs lies in how to balance the competing goals of achieving high accuracy and high efficiency. In this paper we propose two novel types of convolutions, dubbed \emph{Pixel Difference Convolution (PDC) and Binary PDC (Bi-PDC)} which enjoy the following benefits: capturing higher-order local differential information, computationally efficient, and able to be integrated with existing DNNs. With PDC and Bi-PDC, we further present two lightweight deep networks named \emph{Pixel Difference Networks (PiDiNet)} and \emph{Binary PiDiNet (Bi-PiDiNet)} respectively to learn highly efficient yet more accurate representations for visual tasks including edge detection and object recognition. Extensive experiments on popular datasets (BSDS500, ImageNet, LFW, YTF, \emph{etc.}) show that PiDiNet and Bi-PiDiNet achieve the best accuracy-efficiency trade-off. For edge detection, PiDiNet is the first network that can be trained without ImageNet, and can achieve the human-level performance on BSDS500 at 100 FPS and with $<$1M parameters. For object recognition, among existing Binary DNNs, Bi-PiDiNet achieves the best accuracy and a nearly $2\times$ reduction of computational cost on ResNet18. Code available at \href{https://github.com/hellozhuo/pidinet}{https://github.com/hellozhuo/pidinet}.

Figures (19)

• Figure 1: Compared with vanilla (binary) convolution, the proposed (binary) PDC can better capture high-order local differential information that facilitates edge detection and general object recognition. In frequency domain, the high-order information contains more high-frequency components, as illustrated in the FFT2 map averaged over all the feature maps.
• Figure 2: Three instances of pixel difference convolution derived from LBP and ELBP descriptors liu2011sortedliu2012extendedsu2019bird. More instances can be obtained by incorporating other probing strategy of the pixel pairs.
• Figure 3: LBP and its variants with different probing strategies.
• Figure 4: Selection of pixel pairs and convolution in APDC.
• Figure 5: During vanilla BConv, pixels in certain local regions (for example, a pixel patch from the antelope skin) are all binarized to the same value, since the threshold was optimized for the whole input rather than a particular patch area, hence eliminating the valuable micro-structural information. By using a neighboring pixel (the central pixel in this example of Bi-CPDC) as the threshold, in contrast, Bi-PDC effectively preserves such information without any extra parameters introduced.