GMConv: Modulating Effective Receptive Fields for Convolutional Kernels.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) traditionally use fixed square kernels for receptive fields (RFs).
• The effective receptive field (ERF) is crucial as it defines input pixel contribution to output pixels.
• ERFs often follow a Gaussian distribution, a property not fully exploited by standard kernels.

Purpose of the Study:

• To propose a novel convolutional kernel, Gaussian Mask convolutional kernel (GMConv), that refines the receptive field.
• To enhance the performance of CNNs by better approximating the ideal ERF.
• To validate the effectiveness of GMConv across various computer vision tasks.

Main Methods:

• GMConv employs a Gaussian function to create a concentric symmetry mask applied to the convolutional kernel.
• This mask refines the kernel's receptive field, aligning it more closely with the Gaussian distribution of ERFs.
• The approach was evaluated on image classification and object detection benchmarks using standard CNN architectures.

Main Results:

• GMConv demonstrated improved performance compared to standard convolutional kernels across multiple tasks.
• AlexNet and ResNet-50 models incorporating GMConv showed significant accuracy boosts on the ImageNet dataset.
• Specifically, top-1 accuracy increased by 0.98% for AlexNet and 0.85% for ResNet-50.

Conclusions:

• The proposed Gaussian Mask convolutional kernel (GMConv) effectively refines receptive fields in CNNs.
• GMConv offers a simple yet powerful method to enhance CNN performance in computer vision.
• This kernel design represents a promising advancement for deep learning models in image analysis.