Abstract:Target detection on synthetic aperture radar (SAR) imagery is critical to several military and civilian applications. However, the coherent imaging mechanism of SAR will inevitably generate speckle noise, reducing the signal-to-noise ratio of SAR imagery, resulting in the loss of certain important information related to texture, and thus degrading the accuracy of SAR image target detection. In addition, learning irrelevant features, such as natural clutter, increases the false alarm rate. To address these issues, this article proposes a high-accurate and low-latency SAR target detection network (ADERLNet) to enhance the ability of the target detection model to represent valid features from strongly noisy images and irrelevant features. Specifically, the deep target detection architecture incorporates soft thresholding functions as nonlinear transformation layers, which can eliminate noisy and irrelevant features by forcing these interference features to be zeros. Second, given that manually setting the appropriate thresholds in the soft thresholding functions is challenging, the thresholds in ADERLNet are automatically determined to avoid substantial signal processing expertise by integrating specially designed subnetworks as trainable modules. Finally, the classification and regression losses are redesigned to offer the model stronger generalization ability by balancing the positive and negative samples and handling the impact of possible low-quality samples in the training set. Additionally, the structural reparameterization method is employed to balance accuracy and speed. Experimental results on three widely used datasets for SAR target detection, including the SAR ship detection dataset, large-scale multi-class SAR image target detection dataset (MSAR)-1.0, and SAR-Ship-Dataset, reveal that ADERLNet maximally improves average precision (AP) by 2.7%, 3.3%, and 1.8%, respectively, and far outperforms other methods on detection accuracy while maintaining excellent real-time performance, fully highlighting the superiority of ADERLNet. The trained models and source codes will be available at https://github.com/yangyahu-1994/ADERLNet.

SARNas: A Hardware-Aware SAR Target Detection Algorithm Via Multiobjective Neural Architecture Search.

A Zero-Shot NAS Method for SAR Ship Detection Under Polynomial Search Complexity

NAS-ASDet: An Adaptive Design Method for Surface Defect Detection Network using Neural Architecture Search

NAS-FCOS: Efficient Search for Object Detection Architectures

ADERLNet: Adaptive Denoising Enhancement Representation Learning for Low-Latency and High-Accurate Target Detection on SAR Sensors

SM-NAS: Structural-to-Modular Neural Architecture Search for Object Detection

Surrogate-assisted Multi-objective Neural Architecture Search for Real-time Semantic Segmentation

HRLE-SARDet: A Lightweight SAR Target Detection Algorithm Based on Hybrid Representation Learning Enhancement

Advances in neural architecture search

Multi-Objective Hardware Aware Neural Architecture Search using Hardware Cost Diversity

SARGap: A Full-Link General Decoupling Automatic Pruning Algorithm for Deep Learning-Based SAR Target Detectors

Hardware/Software Co-Exploration of Neural Architectures.

Sar Image Target Detection in Complex Environments Based on Improved Visual Attention Algorithm

NAS-FCOS: Fast Neural Architecture Search for Object Detection

Semi-supervised accuracy predictor-based multi-objective neural architecture search

Multi-objective Hardware-aware Neural Architecture Search with Pareto Rank-preserving Surrogate Models

Data Aware Neural Architecture Search

MedNAS: Multi-Scale Training-Free Neural Architecture Search for Medical Image Analysis

A Comprehensive Survey of Neural Architecture Search: Challenges and Solutions

Multi-scale direction-aware SAR object detection network via global information fusion

Learning Efficient Representations for Enhanced Object Detection on Large-scene SAR Images