Volumetric Hadamard Normalization for Synthetic Aperture Sonar Sub-Bottom Target Classification

Image normalization strategies for 3-D synthetic aperture sonar (SAS) is a relatively underexplored area for target classification leveraging convolutional neural networks (CNNs). For 3-D SAS data, underwater sub-bottom targets (i.e., objects that are buried, or partially buried, in submerged sediment) typically exhibit weaker acoustic returns than the frequently cluttered sediment bottom, often disrupting conventional dynamic range compression algorithms (such as log-compression) that use the brightest returns to normalize the rest of the data. This can reduce image contrast of the target relative to clutter features, damaging the performance of the CNN classifier. This work proposes a multiplicative preprocessing layer via the element-wise Hadamard product, attached to the beginning of a CNN architecture, that is trained to emphasize salient target features by enhancing the contrast of the input before 3-D features are extracted via convolutional layers for image classification. The preprocessing layer, referred to as the volumetric Hadamard normalization layer, is trained with a novel template-matching based regularization function, learning to improve the classification performance. Further, by visualizing the output of this trainable layer, we can interpret what the network considers to be salient features of targets. We train and validate our method on measured volumetric SAS data captured by the sediment volume search sonar system, achieving enhanced performance in terms of the area under the precision–recall curve (AUC-PR) score when training on a heavily imbalanced data set relative to using raw and log-compressed versions of the data. We conduct several ablation studies to validate the design choices of our method. This work shows that learning a transformation that normalizes (or enhances the contrast) sub-bottom SAS data cubes on a voxel-level can produce sub-bottom normalized data that achieves better image recognition performance.