How Can a Neural Network Spot Tiny Hidden Objects in Radar Images?

Detecting tiny objects in synthetic aperture radar (SAR) images is a real challenge. Low resolution, arbitrary orientations, and complex backgrounds make standard detectors struggle. To tackle this, researchers designed DVDNet, a convolutional neural network built especially for small SAR targets. DVDNet merges three innovations: grouped-deformable convolution that adjusts the receptive field to various object sizes; a Local Binary Pattern enhancement module that highlights textures and makes camouflaged objects more visible; and a vector decomposition module that regresses oriented bounding boxes using learnable geometric vectors. The system sits inside a two‑stage architecture, preserving the fine details needed for pinpointing tiny objects. Tests on two SAR datasets—HRSID and SSDD—showed impressive results: 90.9 % mAP on HRSID and 87.2 % on SSDD. DVDNet also performed well on a custom SAR ship dataset and on the optical remote‑sensing dataset HRSC2016, confirming its ability to generalize. In practice, by combining deformable sampling, texture enhancement, and vector‑based box representation, DVDNet achieves higher precision and recall, especially for the smallest objects, proving the approach works even in very complex radar scenes.

DVDNet builds on a typical two‑stage detector architecture but introduces three key blocks. The first, Grouped‑Deformable Convolution, splits channels into groups and applies independent deformable convolutions, allowing the kernel to shift sampling points in response to scale and orientation variations. This is especially useful for SAR objects, where sizes can vary dramatically within the same image. The second block, LBP Enhancement Module, computes local binary patterns on intermediate feature maps, emphasizing intensity differences typical of radar textures. Its output is concatenated with the original maps, improving discrimination of low‑contrast or partially hidden objects. The third block, Vector Decomposition Module, represents oriented bounding boxes via two vectors: one for the main direction and one for width. During training, the network learns to predict these vectors directly, avoiding complex angle regressions and reducing rotation errors. The loss combines a classification component (focal loss) with a Smooth‑L1 regression on the vectors. Optimization uses Adam with a cyclical learning rate, and the HRSID and SSDD datasets are split into training/validation with SAR‑specific data augmentation (random rotations, speckle noise addition). Results show a significant mAP boost over Faster‑RCNN and YOLOv5, especially for the small‑object subclass (<10 px). Moreover, generalization tests on HRSC2016 indicate that the texture and vector modules transfer to optical images, suggesting DVDNet captures universal geometric features. Future work includes integrating attention mechanisms and testing the model on higher‑frequency SAR data to assess framework scalability.