DAN: A Deformation-Aware Network for Consecutive Biomedical Image Interpolation

Abstract

The continuity of biological tissue between consecutive biomedical images makes it possible for the video interpolation algorithm, to recover large area defects and tears that are common in biomedical images. However, noise and blur differences, large deformation, and drift between biomedical images, make the task challenging. To address the problem, this paper introduces a deformation-aware network to synthesize each pixel in accordance with the continuity of biological tissue. First, we develop a deformation-aware layer for consecutive biomedical images interpolation that implicitly adopting global perceptual deformation. Second, we present an adaptive style-balance loss to take the style differences of consecutive biomedical images such as blur and noise into consideration. Guided by the deformation-aware module, we synthesize each pixel from a global domain adaptively which further improves the performance of pixel synthesis. Quantitative and qualitative experiments on the benchmark dataset show that the proposed method is superior to the state-of-the-art approaches.

References in corpus (9)

• Adam: A Method for Stochastic Optimization
• Learning Depth from Single Monocular Images Using Deep Convolutional Neural Fields
• FlowNet: Learning Optical Flow with Convolutional Networks
• Deep Ordinal Regression Network for Monocular Depth Estimation
• Single-Image Depth Perception in the Wild
• Semi-Supervised Deep Learning for Monocular Depth Map Prediction
• Video Frame Interpolation via Adaptive Separable Convolution
• Interlaced Sparse Self-Attention for Semantic Segmentation
• Video Frame Interpolation via Adaptive Convolution