Source-linked AI summary
Improved Automatic Target Recognition in Synthetic Aperture Sonar Imagery Using Large Deep Neural Networks
C. J. Moore, Alex Hurt, Jordan Malof
TL;DR
SAS-ATR must recognize targets in imagery despite limited labeled data, while transformers remain less represented than CNNs. This paper systematically compares modern CNNs and transformers across capacity and regularization choices, finding that properly regularized high-capacity DNNs generally achieve stronger SAS-ATR performance.
Problem
SAS-ATR research has limited labeled data and has focused mainly on CNNs, leaving the performance of large transformer and CNN models underexplored.
Method
The paper evaluates six CNN and transformer architectures while systematically varying capacity, pretraining, data augmentation, Drop Path, Weight Decay, and related hyperparameters.
Results
10% AP improvement is reported for both ResNeXt-152 over ResNeXt-50 and Swin-Base over Swin-Tiny; high-capacity models generally outperform low-capacity models.
Takeaways & Limitations
High-capacity DNNs, when properly regularized, represent a promising methodology for SAS-ATR and provide guidance for training such models.
Abstract
from arXiv · showhide
Automatic Target Recognition (ATR) in Synthetic Aperture Sonar (SAS) is a task largely dominated by deep neural networks (DNNs). Most SAS-ATR models use convolutional neural network (CNN) architectures whereas transformer-based architectures have had much less representation in the literature despite being state of the art in general computer vision (CV) research. Additionally, researchers have had mixed results in attempting to overcome challenges presented by a scarcity of labeled training data by using methods such as data augmentation and the use of pretrained weights from a variety of imaging modalities. In this work, we compare the performance of modern CNN and transformer-based DNNs to determine which architecture and training configurations elicit the highest performance in SAS-ATR. We investigate how network size, architecture, pretraining method, data augmentation and other forms of regularization affect SAS-ATR performance with a focus on producing the highest-performing model and providing a roadmap for training state-of-the-art SAS-ATR models.
I. INTRODUCTION
This work asks whether large modern CNN and transformer DNNs, properly regularized, can improve SAS-ATR despite limited labeled data. It evaluates capacity, architecture, and training choices to identify high-performing configurations.
- I. INTRODUCTION: SAS-ATR uses DNNs to process large volumes of seafloor imagery and indicate where target objects are likely to exist.The models map SAS images to target or non-target predictions after training on labeled real-world imagery.
- I. INTRODUCTION: Double Descent motivates testing high-capacity networks because generalization can improve after model capacity exceeds the interpolation threshold.Highly overparameterized networks may perform well even when labeled training data is limited.
- A. Contributions of This Work.: The study compares six architectures, including three CNNs and three transformers, using controlled optimization and evaluation regimes.It also compares high- and low-capacity versions across SAS-ATR data from three geographic locations with leave-one-location-out cross-validation.
- A. Contributions of This Work.: The experiments systematically investigate pretraining, data augmentation, Drop Path, Weight Decay, and their application-specific hyperparameters.Pretraining sources include optical imagery, Synthetic Aperture Radar, and ultrasound imagery; eleven augmentations are evaluated.
- A. Contributions of This Work.: High-capacity models generally outperform low-capacity models, with regularization adaptations substantially improving their SAS-ATR accuracy.The paper presents this as a systematic evaluation aimed at achieving state-of-the-art SAS-ATR accuracy.
II. RELATED WORK
Related work establishes CNNs as the historical SAS-ATR default while introducing growing interest in transformers, transfer learning, augmentation, and regularization. The paper frames these directions around limited labeled SAS data and model-capacity concerns.
- A. SAS-ATR and Neural Network Capacity: Double Descent challenges the assumption that scarce labeled data requires very small networks for good generalization.Earlier SAS work used tiny CNNs partly to address overfitting concerns associated with large capacity.
- B. CNNs and Transformers in SAS-ATR: SAS-ATR research has historically favored CNNs, while transformer and hybrid approaches have received more recent but comparatively limited attention.Vision Transformers have achieved state-of-the-art results on general computer-vision tasks.
- C. Pretraining Neural Networks for SAS: Prior transfer-learning studies report mixed evidence about whether weights from optical imagery improve SAS-ATR performance.Some work favors transfer between SAS sensors, while other research reports benefits from optical-image pretraining.
- D. SAS Data Augmentation: SAS augmentation research distinguishes physical transformations that preserve scene geometry from non-physical transformations.Small affine transformations and spectral information have been reported to improve SAS-ATR performance.
- E. Network Regularization in SAS-ATR: Existing SAS-ATR regularization includes SAS-specific structural priors in loss functions and statistical regularizers on network weights.These approaches incorporate domain knowledge or encourage favorable statistical patterns during learning.
- III. PROBLEM SETTING: The problem is binary classification of real-valued beamformed SAS patches as target or non-target using parameterized functions trained by empirical-loss minimization.The empirical objective minimizes loss between predictions and ground-truth labels over the dataset.
IV. METHODS
This section introduces the experimental methods and the key designer-set hyperparameters used to manipulate training configurations.
- IV. METHODS: The methods section summarizes the experimental procedures and their essential implementation details.It focuses on the hyperparameters varied in the experiments.
- IV. METHODS: Key hyperparameters are parameters set by the designer and manipulated to study their effects in the experiments.The section presents these settings as central experimental controls.
- IV. METHODS: The methods provide the basis for comparing alternative training configurations under controlled experimental variation.This follows from the section’s emphasis on the details and hyperparameters manipulated in the experiments.
A. Deep Neural Networks (DNNs)
DNN effectiveness depends on architecture and capacity, with CNNs and transformers forming the main architecture classes. The paper varies these choices and evaluates training and performance procedures for SAS-ATR.
- A. Deep Neural Networks (DNNs): DNN capacity is defined by total parameters and can be changed through network depth and layer width.
- A. Deep Neural Networks (DNNs): CNNs primarily use convolutional layers, whereas transformers primarily use transformer layers for image processing.
- A. Deep Neural Networks (DNNs): The study compares six architectures—ResNet, ResNeXt, ConvNext, ViT, SWIN, and HiViT—at varied capacities, producing 14 networks.
- B. DNN Training: DNN training uses gradient descent to iteratively adjust parameters and reduce empirical loss over the training data.
- B. DNN Training: The experiments use AdamW, which averages gradients over minibatches before updating parameters and includes minibatch size and weight decay hyperparameters.
- B. DNN Training: Because non-target instances are abundant and often easy to classify, the study uses focal loss with γ = 2.0 and α = 0.25.
- B. DNN Training: Performance is assessed with k-fold cross-validation and reported using Average Precision and Optimal F1.
C. DNN Regularization
The paper treats regularization as a set of techniques for improving generalization in high-capacity DNNs. It investigates pretraining, DropPath, weight decay, and data augmentation for SAS-ATR.
- C. DNN Regularization: The study investigates pretraining, DropPath, weight decay, and data augmentation to mitigate poor generalization in high-capacity DNNs.
- C. DNN Regularization: DropPath randomly removes paths during training, preventing their outputs from contributing to the final representation.
- C. DNN Regularization: Weight decay shrinks network parameters toward zero during training and is implemented through the Adam-W optimizer.
- C. DNN Regularization: Pretraining initializes a network with weights learned from other data, providing prior information about spatial phenomena compared with training from scratch.
- C. DNN Regularization: SAS-ATR augmentation is discussed as either physical, preserving essential SAS properties, or non-physical, which may also benefit performance.
V. MAIN EXPERIMENTS
The main experiments test whether larger DNNs and transformer architectures improve SAS-ATR. They use geographically separated data, controlled model comparisons, and training configurations involving augmentation and regularization.
- V. MAIN EXPERIMENTS: The experiments ask whether high-capacity DNNs outperform smaller models and whether transformers outperform convolutional architectures.
- A. Experimental Design Details: The dataset contains 500,000 beamformed SAS image patches from three geographic locations with distinct oceanic and bathymetric conditions.
- A. Experimental Design Details: Patches are selected from larger seafloor images using an energy-based detector and resized to 256 × 256 pixels for model input.
- A. Experimental Design Details: Each model is evaluated with leave-one-region-out cross-validation to provide a real-world performance assessment.
- A. Experimental Design Details: The classifier training process incorporates data augmentation, network regularization, and pretrained weights as performance-improvement methods.
B. Experimental Results
Across the reported experiments, increasing capacity generally improves performance, and the strongest tested networks are SWIN, ResNeXt, and ConvNeXt. ImageNet pretraining is consistently beneficial relative to alternatives and training from scratch.
- Impact of model capacity: 10% AP separates ResNeXt-152 from ResNeXt-50, while Swin-Base improves 10% AP over Swin-Tiny within their architecture families.
- Impact of model capacity: High-capacity models generally improve over smaller versions, although HiViT and ResNet show modest gains or slight degradation in one case.
- Comparison with Existing SAS-ATR Models: Every proposed high-capacity model outperforms the literature-based ResNet and TinyCNN under the shared training and optimization scheme.
- Best Overall Performing Model?: SWIN, ResNeXt, and ConvNeXt are the most performant tested networks, with their largest variants showing the best performance.
- Impact of Pretraining: ImageNet pretraining is consistently beneficial compared with training from scratch and other pretraining strategies, across both reported performance metrics.
B. Impact of Augmentation
The study evaluates augmentation policies for CNN and transformer SAS-ATR models, finding that optimized augmentation benefits transformers substantially more than convolutional models.
- Impact of augmentation: The augmentation study used ten individual augmentations and searched for a combination, termed a policy, that performed well for SAS-ATR data.Each augmentation had hyperparameters whose settings were measured and optimized.
- Impact of augmentation: SWIN benefited from almost every tested augmentation, whereas ConvNeXt achieved only marginal improvement from optimized hyperparameters.The search optimized augmentation hyperparameters separately for one transformer and one convolutional model because optimal settings may depend on architecture.
- Impact of augmentation: A near 10-point AP gain and 5-point optimal F1-score gain were achieved by SWIN-base using Zoom and Rotation together.These gains were reported in the comparison of augmentation policies for SWIN-base.
- Impact of augmentation: The results provide evidence that transformers may benefit more from augmentation and may perform better with more training data or regularization.This interpretation is described as consistent with broader computer vision literature.
C. Impact of Regularization
Regularization generally improves SAS-ATR performance and helps high-capacity models realize their advantages, with augmentation especially effective for transformers.
- Impact of regularization: Regularization improved the performance of every evaluated network, while smaller networks often received the largest improvements but remained worse than larger counterparts.Regularization comprised weight decay, DropPath, and data augmentation, with hyperparameters optimized for each network.
- Impact of regularization: High-capacity SWIN, ResNeXt, and ConvNeXt variants achieved the best overall performance and substantially outperformed two SAS-ATR baseline models.The study compared six architectures and included two recent models from the SAS-ATR literature.
- Impact of regularization: Regularization was found to be crucial for realizing the advantages of increasing model capacity.This conclusion links the capacity findings to the aggregate regularization experiments.
- Impact of regularization: Data augmentation was highly beneficial for transformers but less beneficial for convolutional models, while DropPath and weight decay were generally beneficial.The conclusion distinguishes augmentation effects from the broader regularization results.
VIII. APPENDIX
The appendix documents implementation resources, augmentation settings, focal-loss behavior, pretrained-weight sources, and reproducibility details.
- Pretrained weights: ImageNet weights came from Torchvision, ultrasound weights from Hugging Face, and SAR weights from an open-source SAR project.The paper provides links to the weight sources for replicability.
- Training infrastructure: Training was conducted with MMPretrain using data-parallel training on four NVIDIA A100X GPUs totaling 320GB of high-bandwidth memory.The hardware configuration supported accelerated training.
- Augmentation implementation: The study used ten augmentations and reports their tested hyperparameters and optimized values for reproducibility.The values were jointly optimized using a grid search and are listed in Table 7.
- Focal Loss: Focal Loss uses α to address class imbalance and γ to reduce the impact of trivial predictions.The paper uses α=0.25, making target-sample losses more influential than non-target losses; high-confidence correct predictions receive reduced impact when γ>1.
E. Best Hyperparameters – Regularization
The appendix reports the highest-performing regularization configurations separately from the performance results.
- Best hyperparameters: Table 9 reports the regularization hyperparameter values found optimal for each network.Table 6 reports the corresponding highest performance, but does not list the optimal values.