An Adaptive Kernels Layer for Deep Neural Networks Based on Spectral Analysis for Image Applications

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Increasing pixel resolution leads to larger image sizes, challenging traditional deep learning models.
• Existing models struggle with location-sensitive tasks like object localization and change detection due to fixed receptive fields.

Purpose of the Study:

• Propose an adaptive convolutional kernels layer (AKL) that dynamically adjusts to image sizes.
• Improve feature spatial resolution without sacrificing the local receptive field (LRF).
• Enhance performance in image applications sensitive to feature locations.

Main Methods:

• Developed an adaptive convolutional kernels layer (AKL) based on Fourier transform and spectral analysis.
• Utilized Monte Carlo simulations to evaluate spectral information coverage across various image sizes.
• Compared AKL performance against Inception and VGG architectures.

Main Results:

• AKL demonstrated consistent spectral information coverage across different image sizes with a variation of ~20%.
• AKL matched Inception v4 in image classification accuracy.
• AKL outperformed Inception v4 in object localization for larger images, achieving up to 30% higher accuracy.

Conclusions:

• The proposed AKL effectively addresses challenges posed by large image sizes in deep learning.
• AKL offers a robust solution for location-sensitive image analysis tasks.
• AKL provides a scalable and accurate approach for image classification and object localization.