A Parameter Estimation-Inspired Convolutional Block for SAR Data

• Cassandra Goldberg University of Pennsylvania
• Jeova Farias Sales Rocha Neto Bowdoin College

Resumo

Convolutional neural networks (CNNs) have been widely used in the context of SAR imagery analysis. However, most of the architectural choices involved in designing CNNs are based on trial-and-error with with limited grounding in formal mathematical principled decision-making. Here, we aim to bridge the gap between traditional moment-based approaches for SAR image processing and modern convolution-based learning algorithms. Specifically, we propose a simplified version of the traditional convolution operation, composed now of an average pooling layer for (first-order) moment computation and a series of 1×1 convolutional layers for learning nonlinear transformations of these moments. In practice, we show that traditional segmentation CNNs incorporating these simplified convolutional layers achieve similar results to their original counterparts while requiring fewer parameters, resulting in reduced computational and memory overhead.

Palavras-chave: Graphics, Image segmentation, Grounding, Neural networks, Statistical learning, Memory management, Decision making, Radar polarimetry, Convolutional neural networks, Synthetic aperture radar, neural network networks, statistical learning, image segmentation