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

Area of Science:

• Computer Vision
• Deep Learning
• Neural Network Architectures

Background:

• Deep feature fusion is crucial for the learning capabilities of Convolutional Neural Networks (CNNs).
• Efficient aggregation strategies and multiscale representations are key areas of advancement in CNNs.
• Existing architectures like DenseNet and ResNet have demonstrated success but have limitations in feature fusion and redundancy.

Purpose of the Study:

• To introduce a novel network architecture, ResDNet, for high-level computer vision tasks.
• To enhance feature fusion and reduce redundancy through densely connected multiscale representations.
• To demonstrate the effectiveness and efficiency of ResDNet compared to existing state-of-the-art models.

Main Methods:

• Developed ResDNet, a backbone architecture using sequential ResDNet modules with Sliding Dense Blocks (SDBs).
• Employed densely connected feature fusion to provide multiscale representations within a residual network framework.
• Evaluated ResDNet on CIFAR-10, CIFAR-100, ImageNet classification benchmarks, and COCO object detection dataset using RetinaNet.

Main Results:

• ResDNet outperforms DenseNet on CIFAR-100 with significantly less computation.
• ResDNet-B-129 achieved superior top-1 accuracy over ResNet-50 and DenseNet-201 on ImageNet with comparable complexity.
• ResDNet-B-129 improved mean Average Precision (mAP) on the COCO dataset when implemented with RetinaNet, outperforming ResNet-50.

Conclusions:

• ResDNet offers enhanced feature fusion and reduced redundancy compared to DenseNet through shallower densely connected architectures.
• The proposed ResDNet architecture demonstrates strong performance and efficiency across various computer vision tasks, including classification and detection.
• ResDNet represents a significant advancement in neural network design for computer vision, achieving state-of-the-art results.