Project Summary

• Developed PyNASDet, a novel framework integrating the hardware-aware evolutionary Neural Architecture Search engine PyNAS with the MMDetection object detection pipeline, for the automated design of efficient SAR ship detection models.
• The framework introduces the NASEncoder module as the core interface between the NAS search space and the detector: it dynamically reconstructs backbone networks from compact architecture strings generated by the genetic algorithm, enabling full automation of the architecture-to-detector pipeline.
• An evolutionary search campaign was conducted on the HRSID dataset over 10 generations. The best-discovered model achieved an accuracy–efficiency trade-off that is more than one order of magnitude faster than conventional reference detectors (Faster R-CNN, Mask R-CNN), with only a limited reduction in detection performance.
• PyNASDet establishes a reusable methodological foundation for hardware-aware detector design in resource-constrained Earth Observation scenarios, with a natural extension path toward range-compressed SAR data and edge/onboard deployment.

Development Tools

• PyNAS: hardware-aware evolutionary NAS framework developed at ESA Φ-lab, providing the genetic algorithm search engine, the search space definition, and the Hardware-Awareness Block (HAB) for device-level performance profiling
• MMDetection (MMDet): open-source object detection toolbox from the OpenMMLab ecosystem, built on PyTorch. Used as the execution engine for detector instantiation, training, validation, testing, and COCO-metric evaluation.
• HRSID (High-Resolution SAR Images Dataset): SAR ship detection benchmark containing 5,604 image chips (800×800 px) from 136 panoramic SAR scenes, with 16,951 annotated ship instances in COCO format.
• PyTorch 2.6.0 with CUDA 12.4.
• NVIDIA GeForce RTX 3090 (24 GB VRAM): GPU used for all training and evaluation experiments during the visiting period at ESA Φ-lab (ESRIN).

Development Outputs

• Project Repo: https://github.com/valeriottad12-sudo/PyNASDet
• Master's thesis: "PyNASDet: Neural Architecture Search for Resource-Constrained Satellite Object Detection" — Valeria Biscardi, Università degli Studi di Napoli Federico II, Dipartimento di Ingegneria Industriale, Academic Year 2024–2025. Supervisors: Prof. Alfredo Renga, Dr. Roberto Del Prete. The thesis documents the full methodological development, experimental campaign, and results of the PyNASDet framework.

Project Description

Context and Motivation

Ship detection in SAR imagery is a well-established Earth Observation task, where deep learning detectors have substantially advanced the state of the art. However, these models rely on heavyweight backbones that make onboard deployment impractical under the memory, power, and latency constraints typical of spaceborne platforms. NAS has emerged as a more principled alternative to post hoc compression techniques such as pruning, quantization, and knowledge distillation, enabling the automated discovery of architectures natively optimized for both task performance and hardware efficiency. The visiting period at ESA Φ-lab aimed to extend PyNAS, a hardware-aware evolutionary NAS framework developed at Φ-lab, to the object detection domain, specifically targeting SAR ship detection on the HRSID benchmark.

The PyNASDet Framework

The core contribution of the collaboration is PyNASDet, a unified pipeline that integrates PyNAS with the MMDetection (MMDet) framework. PyNAS provides an evolutionary search engine based on a genetic algorithm that explores a structured space of candidate backbones parametrized by layer types, kernel sizes, channel scaling factors, activation functions, and downsampling strategies. MMDet provides a mature, modular infrastructure for training and evaluating object detectors, supporting standard detection metrics and a wide range of architectures.

The central technical challenge was to bridge two independent frameworks that use different representations of neural networks. This was solved through the NASEncoder module. Each candidate architecture generated by PyNAS is encoded as a compact string of structured tokens describing the ordered sequence of layers and their hyperparameters. The NASEncoder reads this string, parses it into a sequence of layer specifications, and dynamically reconstructs the corresponding PyTorch backbone. The downstream components of the detector — the NAS_Neck (an adaptive Feature Pyramid Network) and the RADVFCFNet detection head — automatically adapt to the channel dimensions of the reconstructed backbone, enabling the construction of a complete, trainable detector without any manual intervention.

The architecture transfer between the two frameworks is implemented via a file-based interface: the NAS controller writes the current candidate's architecture string to a shared file, which MMDet then reads at detector instantiation. This design keeps the two systems decoupled while ensuring a reproducible and auditable interface.

The fitness function adopted by PyNASDet combines COCO-based detection accuracy (mAP_50) with measured inference throughput (FPS) via a hardware-aware formulation, explicitly favoring architectures that jointly maximize detection performance and runtime efficiency. A dynamic gating strategy further penalizes architectures with insufficient detection capability in later generations, making the evolutionary search progressively more selective.

Experimental Campaign and Results

The experimental campaign was conducted on the High-Resolution SAR Images Dataset (HRSID), a benchmark specifically designed for SAR ship detection containing 5,604 image chips of size 800×800 pixels derived from 136 panoramic scenes acquired by Sentinel-1B, TerraSAR-X, and TanDEM-X. The dataset was split into 2,914 training images, 729 validation images, and 1,961 test images.

The NAS search was performed over 10 generations, with a population of 10 individuals per generation, yielding 73 effectively trained and evaluated candidate architectures (duplicates are automatically skipped). A progressive training schedule was adopted: early generations were trained for 10 epochs, intermediate generations for 15 epochs, and the final generations for up to 25 epochs. All experiments were carried out on an NVIDIA GeForce RTX 3090 GPU with 24 GB of VRAM.

The best PyNASDet model, from Generation 10, achieved mAP_50 = 0.783 at an inference speed of 338 FPS. This result corresponds to an order-of-magnitude increase in throughput compared to conventional ResNet-50-based detectors evaluated on the same benchmark (Mask R-CNN: 12 FPS, Cascade Mask R-CNN: 10.3 FPS), with only a limited reduction in detection accuracy (mAP_50 for the reference detectors ranges from 0.847 to 0.893). Notably, the mAP_S value of the best PyNASDet model (0.538) is comparable to those of the reference detectors, suggesting that the NAS-discovered architecture preserves small-object detection capability despite operating at a much higher speed.

The generational analysis confirms a clear convergence behavior: the best mAP_50 per generation increased from 73.1% in Generation 1 to 78.3% in Generation 10, with the top-5 models all concentrated in the final generations. The worst-fitness curve remains irregular throughout the search, indicating that the evolutionary process maintains sufficient exploratory diversity while exploiting high-performing regions of the design space.

Table 1- Top 5 PyNASDet models ranked by fitness

Figure 1- PyNASDet models' performances across generations