Using a model of human visual perception to improve deep learning.

Area of Science:

• Computer Vision
• Machine Learning
• Cognitive Science

Background:

• Deep learning models excel at image classification but can be brittle when encountering occluded objects.
• Human visual perception effectively handles object recognition despite partial occlusion, suggesting representational advantages.

Purpose of the Study:

• To investigate the impact of occlusion on deep learning image classification performance.
• To develop and evaluate a human-inspired image pre-processing technique to mitigate occlusion effects.

Main Methods:

• A deep learning system was trained on the MNIST digit dataset.
• Occlusion was introduced to test system robustness.
• A novel segmentation and interpolation algorithm, inspired by human visual processing, was developed to reconstruct occluded image regions.
• The reconstruction algorithm was used to pre-process occluded images before classification.

Main Results:

• Occluding objects significantly impaired the deep learning system's classification accuracy.
• Pre-processing occluded images with the human-inspired reconstruction algorithm led to improved training and classification performance.
• The method demonstrated enhanced robustness against partial occlusions.

Conclusions:

• Human visual system's representational format may offer advantages for robust image classification.
• A human-inspired image reconstruction pre-processing step can enhance deep learning model performance on occluded data.
• This approach offers a promising direction for improving the robustness of artificial visual systems.