LD-Det: Lightweight Ship Target Detection Method in SAR Images via Dual Domain Feature Fusion

Ship detection technology represents a significant research focus within the application domain of synthetic aperture radar. Among all the detection methods, the deep learning method stands out for its high accuracy and high efficiency. However, large-scale deep learning algorithm training requires huge computing power support and large equipment to process, which is not suitable for real-time detection on edge platforms. Therefore, to achieve fast data transmission and little computation complexity, the design of lightweight computing models becomes a research hot point. In order to conquer the difficulties of the high complexity of the existing deep learning model and the balance between efficiency and high accuracy, this paper proposes a lightweight dual-domain feature fusion detection model (LD-Det) for ship target detection. This model designs three effective modules, including the following: (1) a wavelet transform method for image compression and the frequency domain feature extraction; (2) a lightweight partial convolutional module for channel feature extraction; and (3) an improved multidimensional attention module to realize the weight assignment of different dimensional features. Additionally, we propose a hybrid IoU loss function specifically designed to enhance the detection of small objects, improving localization accuracy and robustness. Then, we introduce these modules into the Yolov8 detection algorithm for implementation. The experiments are designed to verify LD-Det’s effectiveness. Compared with other algorithm models, LD-Det can not only achieve lighter weight but also take into account the precision of ship target detection. The experimental results from the SSDD dataset demonstrate that the proposed LD-Det model improves precision (P) by 1.4 percentage points while reducing the number of model parameters by 20% compared to the baseline. LD-Det effectively balances lightweight efficiency and detection accuracy, making it highly advantageous for deployment on edge platforms compared to other models.