Non-cooperative bistatic radar has a low signal-to-noise ratio in the echo due to its special detection method. In particular, the detection between frames in the radar scanning cycle for maritime moving targets is not stable, which will bring great difficulties for subsequent target tracking. The low threshold Constant False Alarm Rate(CFAR) detector is employed to match the detection results of radar range-Doppler dimension and range-azimuth dimension to obtain the corresponding mask map, and the potential moving targets are found. Then, a Double Backbone-YOLO(DB-YOLO) that fuses multi-dimensional feature information is proposed. The network adopts a dual-trunk structure, extracts the features of the moving target mask map and the same-scale P-display map under its mapping, and uses a deep separable convolution module to reduce the model parameters of the network. Finally, the comparison experiments with Faster RCNN, YOLOv5 and its common variant YOLOv5-ConvNeXt show that DB-YOLO effectively improves the target detection performance and ensures the inference speed, which lays a foundation for target tracking of noncooperative bistatic radar