SEFEPNet: Scale Expansion and Feature Enhancement Pyramid Network for SAR Aircraft Detection With Small Sample Dataset

Abstract

Aircraft detection in Synthetic Aperture Radar (SAR) images is still a challenging research task because of the insufficient public data, the difficulty of multi-scale target detection, and the complexity of background interference. In this paper, we construct a public SAR Aircraft Detection Dataset (SADD) with complex background and interference objects to facilitate the research in SAR aircraft detection. Then, we propose the Scale Expansion and Feature Enhancement Pyramid Network (SEFEPNet) as the SADD baseline. It uses a four-scale fusion method to combine the shallow position information with the deep semantic information, effectively adapting to the multiscale target detection in SAR images, significantly improving the detection effect of small targets. The Feature Enhancement Pyramid (FEP) structure is connected behind the backbone network to weaken the background texture and highlight the target to achieve feature enhancement, improving the ability to extract target features in complex backgrounds. Finally, to further improve the detection performance of the small-scale SAR aircraft dataset, we propose a domain adaptive transfer learning method. Experiments on SADD show that this method can significantly improve the recall rate and F1 score. At the same time, we find that the transfer effect of using homologous but different types of targets as the source domain is better than those of heterologous but same types of targets in SAR aircraft detection, which is instructive for future research.

YOLO-SAD: An Efficient SAR Aircraft Detection Network

“…The CBS(k, s) module includes a convolutional layer with a kernel size of n and stride m, followed by batch normalization and SiLU [47] activation for nonlinear transformations. The MP module merges CBS (3,2) and pooling layers to achieve 8×, 16×, and 32× down-sampling, generating feature maps of the 1/8, 1/16 and 1/32 input image's size, respectively. This process enlarges the receptive field and decreases feature map dimensions.…”

“…For example, the Salient Fusion Module (SFM) integrates high-resolution details with semantic features to enhance feature discrimination in complex scenes [2]. Zhang et al introduced the SAR Aircraft Detection Dataset (SADD) and the SEFEPNet, a deep learning model leveraging domain adaptive transfer learning for better detection [3]. Additionally, Han et al's Low-Level Semantic Enhancement Module (LSEM) focuses on enhancing scattered features in SAR imagery [4].…”

“…The SADD is a publicly available dataset for SAR aircraft detection, comprising 2966 224 × 224 pixel images [3]. Noting that the dataset had been augmented for diversity, we conducted a fresh split to prevent duplication between training and validation sets.…”

“…Combining both channel and spatial attention mechanisms before the ELAN module achieved the highest overall accuracy, with a 56.6% AP(0.5:0.95) and a 92.2% AP(0.5) when using kernel sizes of [5,7,5,3]. Despite no change in model size and complexity, detection speed decreased by 0.6 to 1.8 FPS compared to using only channel attention.…”

Unsupervised aircraft detection in SAR images with image-level domain adaption from optical images

“…The summarized details of the dataset are given in Table 1. The target SAR domain adopts the SAR aircraft detection dataset (SADD) [8] and the source optical domain collects optical images from the public aerial image dataset DOTA [9] . In SADD, there are 2966 SAR chips with an image size of 224×224 collected from the TerraSAR-X satellite.…”