Abstract:In this work, we present a simple, highly efficient and modularized Dual Path Network (DPN) for image classification which presents a new topology of connection paths internally. By revealing the equivalence of the state-of-the-art Residual Network (ResNet) and Densely Convolutional Network (DenseNet) within the HORNN framework, we find that ResNet enables feature re-usage while DenseNet enables new features exploration which are both important for learning good representations. To enjoy the benefits from both path topologies, our proposed Dual Path Network shares common features while maintaining the flexibility to explore new features through dual path architectures. Extensive experiments on three benchmark datasets, ImagNet-1k, Places365 and PASCAL VOC, clearly demonstrate superior performance of the proposed DPN over state-of-the-arts. In particular, on the ImagNet-1k dataset, a shallow DPN surpasses the best ResNeXt-101(64x4d) with 26% smaller model size, 25% less computational cost and 8% lower memory consumption, and a deeper DPN (DPN-131) further pushes the state-of-the-art single model performance with about 2 times faster training speed. Experiments on the Places365 large-scale scene dataset, PASCAL VOC detection dataset, and PASCAL VOC segmentation dataset also demonstrate its consistently better performance than DenseNet, ResNet and the latest ResNeXt model over various applications.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is to design a new deep neural network architecture to further improve the performance of image classification tasks and overcome the limitations of existing architectures (such as ResNet and DenseNet). Specifically: 1. **Feature Reuse and Exploration**: - ResNet realizes feature reuse through the residual path. - DenseNet explores new features through the densely connected path. 2. **Combining the Advantages of Both**: - The paper proposes a Dual Path Network (DPN), aiming to inherit the advantages of ResNet and DenseNet simultaneously: it can effectively reuse features and also flexibly explore new features. 3. **Improving Efficiency**: - By introducing the dual - path architecture, DPN is superior to existing advanced models (such as ResNeXt) in terms of parameter efficiency, computational cost, and memory consumption. 4. **Verifying Wide Applicability**: - The paper not only demonstrates the superior performance of DPN on image classification tasks but also conducts experiments on tasks such as scene classification, object detection, and semantic segmentation, proving its wide applicability and consistent advantages. ### Specific Problems and Solutions - **Problem**: ResNet and DenseNet each have their own advantages and disadvantages. ResNet is good at feature reuse but has difficulty exploring new features; DenseNet is the opposite. Although it can explore new features, it has a high degree of redundancy. - **Solution**: By introducing the dual - path architecture, DPN can maintain the ability to explore new features while sharing features, thus achieving better performance and higher efficiency. ### Experimental Results - **ImageNet - 1k Dataset**: - The shallow DPN (DPN - 92) reduces the top - 1 error rate by 0.5% compared to ResNeXt - 101(32×4d) while reducing FLOPs. - The deep DPN (DPN - 131) outperforms the single - model performance of Very Deep PolyNet with a smaller model size. - **Places365 - Standard Dataset**: - DPN - 92 achieves a higher validation accuracy than other methods with fewer parameters. - **PASCAL VOC Dataset**: - On object detection and semantic segmentation tasks, DPN also performs well, further verifying its wide applicability in different tasks. In conclusion, by proposing DPN, this paper successfully solves the trade - off problem between feature reuse and exploration in existing deep neural network architectures and shows significant performance improvement and higher efficiency in multiple tasks.

Dual Path Networks

DPNet: Dual-Path Network for Real-Time Object Detection With Lightweight Attention

Dynamic Multi-path Neural Network

A Dual-Path Small Convolution Network for Hyperspectral Image Classification

Mixed Link Networks.

DPSSD: Dual-Path Single-Shot Detector

Double reuses based residual network

Dual-path Processing Network for High-resolution Salient Object Detection

Dual-Path Feature Aware Network for Remote Sensing Image Semantic Segmentation

DPTNet: A Dual-Path Transformer Architecture for Scene Text Detection

A Dual-Path and Lightweight Convolutional Neural Network for High-Resolution Aerial Image Segmentation

HyperDense-Net: A hyper-densely connected CNN for multi-modal image segmentation

DPNet: Dual-Pyramid Semantic Segmentation Network Based on Improved Deeplabv3 Plus

Dual-Path Feature Fusion Network for Semantic Segmentation of Remote Sensing Images

HS-ResNet: Hierarchical-Split Block on Convolutional Neural Network

Deep Dual-Resolution Networks for Real-Time and Accurate Semantic Segmentation of Traffic Scenes

ResDNet: Efficient Dense Multi-Scale Representations with Residual Learning for High-Level Vision Tasks

DPED: Bio-inspired dual-pathway network for edge detection

Dual-channel deep graph convolutional neural networks

DRINet: A Dual-Representation Iterative Learning Network for Point Cloud Segmentation

Real-Time Dense Semantic Labeling with Dual-Path Framework for High-Resolution Remote Sensing Image