Learning Lightweight Dynamic Kernels With Attention Inside via Local-Global Context Fusion.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Traditional convolutional neural networks (CNNs) utilize fixed kernels across all input positions, potentially limiting feature representation.
• Dynamic convolution methods generate input-specific kernels to enhance model capacity but often result in a significant increase in parameters.

Purpose of the Study:

• To propose a lightweight dynamic convolution method that strengthens CNNs without a substantial rise in parameters and computational cost.
• To improve feature extraction capabilities by learning attention within convolutional kernels.

Main Methods:

• Introduced a novel dynamic convolution approach that learns attention within convolutional kernels, adjusting weights during feature aggregation.
• Employed an auxiliary network to dynamically modify kernel weights, considering local and global contexts.
• Integrated this method with various CNN backbones for image classification and object detection.

Main Results:

• Achieved remarkable improvements in image classification accuracy on CIFAR and ImageNet datasets.
• Demonstrated effectiveness in object detection tasks.
• Showcased significant performance gains with only a minor increase in model parameters and multiply-adds.

Conclusions:

• The proposed lightweight dynamic convolution method effectively enhances CNN performance.
• This approach offers a parameter-efficient way to boost representation ability in deep learning models.
• The method's versatility is confirmed across different computer vision tasks and network architectures.