*4.1. Quantitative Results*

Table 4 shows the quantitative results of Lite-YOLOv5 on the LS-SSDD-v1.0 dataset. In Table 4, we can see the detection performance comparison with the raw YOLOv5. The ablation studies about the influence of each proposed module will be introduced in detail in Section 5 by the means of each installation and removal.

From Table 4, one can conclude that:


**Table 4.** The performance comparison with the raw YOLOv5. P: Precision, the higher the better; R: recall, the higher the better; AP: average precision, the higher the better; *F1*: F1-score, a main evaluation index, the higher the better; FLOPs: floating point operations, refer to model complexity; model volume: refers to size of model weight; T: the running time of a large-scale image, refers to detection speed (tested on the Jetson TX2).


Table 5 shows the performance comparisons of Lite-YOLOv5 with eight other state-ofthe-art detectors. In Table 5, we mainly select the Libra R-CNN [56], Faster R-CNN [57], EfficientDet [58], free anchor [59], FoveaBox [60], RetinaNet [61], SSD-512 [62], and YOLOv5 [27] for comparison. They were all trained on the LS-SSDD-v1.0 dataset with loading ImageNet pre-training weights. Their implementations were also kept basically the same as in the original report. In addition, it should be emphasized that there is no end-to-end on-board SAR ship detector. Thus, we selected the mainstream two-stage detector (i.e., Libra R- CNN, Faster R-CNN) and single-stage detectors (i.e., EfficientDet, free anchor, FoveaBox, RetinaNet, SSD-512, and YOLOv5) in the CV community for comparison.

From Table 5, the following conclusions can be drawn:


**Table 5.** The performance comparisons of Lite-YOLOv5 with eight other state-of-the-art detectors. The best model is marked in bold. Parameter Size refers to the model complexity. (tested on the RTX3090).

