Feature Map Retargeting to Classify Biomedical Journal Figures.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are pivotal in image analysis.
• Feature map retargeting is crucial for optimizing CNN performance.
• Existing methods face challenges with computational cost and memory at deeper layers.

Purpose of the Study:

• To propose a novel "Retarget" layer for enhancing feature map retargeting in CNNs.
• To improve the performance of CNNs in document subfigure classification tasks.
• To address computational and memory inefficiencies in deep CNNs.

Main Methods:

• Introduced a "Retarget" layer that densely samples feature map channels using a spatial attention regressor.
• Replaced convolutional components with depthwise convolutions in a saliency-based distortion layer.
• Integrated the Retarget layer into pre-trained CNNs (DenseNet121) and applied it to document subfigure classification datasets (ImageCLEF2013, 2015, 2016).
• Experimented with nearest neighbor interpolation to approximate spatial sampling for efficiency.

Main Results:

• The redesigned DenseNet121 with the Retarget layer achieved state-of-the-art results on the visual category of the document subfigure classification task without data augmentation.
• The proposed layer demonstrated consistent improvement over baseline and other state-of-the-art attention models.
• Approximation of spatial sampling showed efficiency gains while maintaining performance improvements.

Conclusions:

• The novel Retarget layer effectively enhances feature map retargeting in CNNs.
• The method achieves state-of-the-art performance in document subfigure classification.
• The approach offers a computationally efficient solution for deep CNNs.