Diversity-Learning Block: Conquer Feature Homogenization of Multibranch.

Area of Science:

• Computer Vision
• Deep Learning Architectures
• Neural Network Optimization

Background:

• Visual Geometry Group (VGG)-style Convolutional Neural Networks (ConvNets) are NPU-friendly but lack accuracy.
• Existing reparameterization methods struggle with parallel branch homogenization and fixed kernel shapes, limiting spatial perception.

Purpose of the Study:

• To propose a novel diversity-learning (DL) block to enhance feature learning and enrich the feature space.
• To develop a lightweight and efficient ConvNet, termed DLNet, for improved performance on NPU hardware.

Main Methods:

• Introduced a diversity-learning (DL) block to create the DLNet architecture.
• Incorporated groupwise operations to balance computational cost (FLOPs) and accuracy.
• Developed a lightweight version, DLGNet, for efficient NPU deployment.

Main Results:

• DLGNet achieved comparable performance to state-of-the-art networks across various computer vision tasks including image classification, object detection, and semantic segmentation.
• Demonstrated significant speed improvements: 183% faster than GhostNet and over 600% faster than MobileNetV3 on NPUs with similar accuracy.

Conclusions:

• The proposed DL block and DLGNet effectively enrich feature representation and improve efficiency.
• DLGNet offers a compelling balance of accuracy and speed for NPU-accelerated computer vision applications.