Spatial oblivion channel attention targeting intra-class diversity feature learning.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) excel at image classification but struggle with intra-class sample diversity.
• Diverse features can disturb CNNs, limiting effective feature learning and overall performance.

Purpose of the Study:

• To introduce a novel Spatial Oblivion Channel Attention (SOCA) module for improved intra-class diversity feature learning.
• To enhance the ability of CNNs to learn detailed image features without redundancy.

Main Methods:

• Proposed SOCA performs spatial structure oblivion and progressive regularization on each channel post-convolution.
• SOCA reassigns channel weights progressively to focus on detailed features in an orderly manner.
• Implemented and tested SOCA on standard datasets (CIFAR-10/100) and specialized garbage datasets.

Main Results:

• SOCA demonstrated average accuracy improvements across various CNN architectures including SqueezeNet, MobileNet, BN-VGG-19, Inception, and ResNet-50.
• Specific average accuracy gains ranged from 1.18% to 2.27% depending on the network architecture.
• Class activation maps confirmed SOCA activates more local detail feature regions, validating its interpretability.

Conclusions:

• The proposed SOCA module effectively addresses intra-class diversity challenges in CNN-based image classification.
• SOCA enhances feature learning by focusing on regional details and preventing feature redundancy, leading to significant accuracy improvements.