DiCENet: Dimension-Wise Convolutions for Efficient Networks.

Area of Science:

• Computer Vision
• Deep Learning Architectures

Background:

• Convolutional Neural Networks (CNNs) are fundamental to computer vision.
• Efficient CNN architectures are crucial for resource-constrained devices.
• Existing efficient convolution methods like depth-wise separable convolutions have limitations.

Purpose of the Study:

• Introduce a novel and generic convolutional unit, the DiCE unit.
• Enhance the efficiency and performance of deep learning models.
• Improve spatial and channel-wise information encoding.

Main Methods:

• The DiCE unit employs dimension-wise convolutions for light-weight filtering across input tensor dimensions.
• Dimension-wise fusion efficiently combines these representations.
• DiCE units are integrated into the DiCENet model architecture.

Main Results:

• DiCE units demonstrate significant improvements over depth-wise separable convolutions across various architectures.
• The DiCENet model achieves state-of-the-art results in image classification, object detection, and semantic segmentation.
• DiCENet shows 2-4% higher accuracy on ImageNet compared to models like MobileNetv2 and ShuffleNetv2.
• DiCENet exhibits superior generalization for resource-constrained tasks, outperforming efficient networks like MobileNetv3 and MixNet.

Conclusions:

• The DiCE unit offers a simple yet effective method to enhance CNNs.
• DiCENet provides a highly efficient and performant architecture for diverse computer vision applications.
• The proposed approach is suitable for real-world applications on devices with limited computational resources.