Synthetic aperture radar (SAR) images are characterized by their single channel, low resolution, and low signal-to-noise ratio, whereas target detection methods based on visible image design lack the corresponding optimization. Moreover, Many ship detection tasks need to be run on resource-constrained embedded devices, which poses new challenges to the performance and model volume of detection networks. Hence, this study introduces a lightweight SAR ship detection approach that enhances contour information to tackle these challenges. Initially, anisotropic diffusion filtering and a four-directional Sobel operator are applied to expand the single-channel SAR image to three channels for network learning. Then, drawing inspiration from the lightweight feature extraction network FasterNet and the non-local attention mechanism, an innovative lightweight backbone feature extraction network is designed. This network adeptly models the long-distance contextual relationships of features, achieves multiscale feature fusion, and reduces the parameter count of the detection model without compromising detection accuracy. Algorithm evaluations on public data sets, satellite ship detection data sets (SSDD), and high-resolution SAR image data sets (HRSID) demonstrate that the proposed network excels not only in reducing model size and complexity but also in maintaining high detection accuracy.