Histogram Layers for Synthetic Aperture Sonar Imagery

Joshua Peeples, Alina Zare, Jeffrey Dale, James Keller

Abstract

Traditional Convolutional Neural Networks (CNNs) often exhibit a "shape bias," prioritizing spatial structures over the statistical distributions of pixel intensities. In the context of Synthetic Aperture Sonar (SAS) imagery, where seafloor textures like sand ripples and rocky terrain are defined by complex statistical patterns, this bias can limit classification performance. This paper explores the integration of learnable histogram layers into deep learning architectures to capture these statistical textures. By utilizing a differentiable radial basis function (RBF) kernel for binning, the proposed models achieve significant performance gains over standard CNNs in both synthetic and real-world SAS datasets.

Introduction

Synthetic Aperture Sonar (SAS) provides high-resolution acoustic imagery essential for underwater environmental mapping and target recognition. While deep learning has significantly advanced image analysis, standard CNNs are primarily designed to identify structural arrangements. In underwater domains, environmental factors such as speckle noise and varying seafloor gradients can alter the appearance of textures. Traditional handcrafted statistical features are robust to these changes but require manual tuning. This research bridges the gap by introducing a learnable histogram layer that allows neural networks to automatically extract and utilize statistical texture features alongside spatial ones.

Methodology

The Learnable Histogram Layer

The core innovation is the implementation of a local, radial basis function (RBF) histogram layer. Unlike traditional histograms with fixed bins, this layer treats bin centers and widths as learnable parameters optimized during backpropagation. The RBF kernel provides a "soft" assignment of pixel values to bins, ensuring the operation is differentiable and suitable for end-to-end training within a neural network.

Architectural Configurations

The researchers evaluated two primary integration strategies using a ResNet18 backbone:

  1. Series Configuration: The histogram layer is placed at the end of the feature extraction pipeline.
  2. Parallel Configuration (HistRes): A histogram layer runs in parallel with the global average pooling layer. This allows the model to concatenate structural features (from the CNN) with statistical features (from the histogram layer), providing a comprehensive representation of the seafloor.

Experimental Setup

Datasets

Two distinct datasets were used to validate the approach:

  • Statistical PISAS Dataset: A synthetic collection of 14,208 images across 10 classes (e.g., sand ripples, rocks). These images were modified with specific statistical distributions (multinomial, binomial, and constant) to isolate the models' ability to recognize pixel-level statistics.
  • Multi-site Dataset: A real-world dataset comprising 118 high-resolution SAS images from diverse seafloor locations, categorized into four classes: craters, flat seafloor, rocky areas, and sand ripples.

Results and Discussion

Synthetic Performance and Statistical Robustness

Experiments on the PISAS dataset highlighted the limitations of standard CNNs. While CNNs performed well on structural changes, they struggled when only the statistical distribution of pixels changed. The Histogram-based model achieved 81.11% accuracy, significantly outperforming the standard CNN's 55.65%. When evaluating purely statistical labels, the Histogram layer reached an impressive 99.90% accuracy, compared to the CNN's 57.37%.

Real-World Application

In real-world scenarios, the Parallel HistRes model demonstrated superior generalization. It achieved a top accuracy of 90.33% on the multi-site dataset, surpassing the baseline ResNet18 accuracy of 84.66%. Visualization via t-SNE projections indicated that the histogram layer naturally clusters images with similar pixel intensities, suggesting it learns a universal statistical representation of seafloor types that transcends specific geographic locations.

Conclusion

Integrating learnable histogram layers into deep learning models effectively addresses the "shape bias" of standard CNNs in SAS imagery. By combining spatial awareness with statistical texture extraction, the Parallel HistRes architecture provides a robust solution for seafloor classification. Future work will explore the application of these layers to semantic segmentation tasks and investigate the performance of different differentiable binning functions and deeper backbone architectures.

Summary updated

BibTeX
@inproceedings{peeples2022histogram,
  author = {Joshua Peeples and Alina Zare and Jeffrey Dale and James Keller},
  title = {{Histogram Layers for Synthetic Aperture Sonar Imagery}},
  booktitle = {ICMLA},
  year = {2022},
  doi = {10.1109/icmla55696.2022.00032},
}
Figure