A Long Skip Connection for Enhanced Color Selectivity in CNN Architectures

Area of Science:

• Computer Vision
• Neuroscience-inspired AI
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) exhibit limited color selectivity in natural image datasets.
• CNNs are inspired by the hierarchical structure of the visual cortex.
• Existing CNN architectures may not fully leverage low-level visual information in deeper layers.

Purpose of the Study:

• To enhance the color selectivity and classification accuracy of CNNs.
• To investigate the impact of long skip connections (LSCs) on CNN performance.
• To improve the integration of low-level and high-level features within CNNs.

Main Methods:

• Proposed adding a long skip connection (LSC) between the first and last blocks of the CNN feature extraction stage.
• Applied LSC strategy to established CNN architectures.
• Quantitatively and qualitatively analyzed improvements in accuracy and color selectivity.
• Developed a novel color representation method for feature map visualization.

Main Results:

• LSCs generally improve CNN classification accuracy.
• The proposed LSC demonstrated superior performance over standard skip connections.
• Enhanced color selectivity was observed in CNNs augmented with LSCs.
• The new color representation facilitated manageable qualitative analysis of feature maps.

Conclusions:

• Long skip connections are effective in boosting CNN accuracy and color selectivity.
• Integrating shallow and deep layer information via LSCs is beneficial for processing natural scenes.
• The proposed LSC strategy offers a promising direction for developing more perceptually aligned AI vision systems.