**5. Ablation Study**

In this section, we will introduce the ablation studies to show the influence of each proposed module by the means of each installation and removal.

### *5.1. Ablation Study on the L-CSP Module*

Table 6 shows the ablation study of Lite-YOLOv5 with and without the L-CSP module. In Table 6, "✘" means Lite-YOLOv5 without the L-CSP module (while keeping the other five modules), while "✔" means Lite-YOLOv5 with the L-CSP module (i.e., our proposed detector). From Table 6, one can find that the L-CSP module can guarantee a lighter architecture with 4.44 G FLOPs and a 2.38 M model volume (~45.6% decrease of FLOPs and ~7.0% decline of model volume when compared with Lite-YOLOv5 without the L-CSP module), which confirms that the L-CSP module can offers a model with greatly reduced computation. In addition, there is only a slight decrease to the overall detection performance. Thus, the L-CSP module can realize a model of sharply reduced computation with only a slight accuracy loss, which confirms its superior cost-effectiveness in lightweight network design.


**Table 6.** The ablation study of Lite-YOLOv5 with and without L-CSP module.

### *5.2. Ablation Study on Network Pruning*

We conducted two ablation experiments on network pruning. Experiment 1 in Section 5.2.1 shows the effectiveness of network pruning in Lite-YOLOv5. Experiment 2 in Section 5.2.2 shows the effectiveness of channel-wise pruning in network pruning.

### 5.2.1. Experiment 1: Effectiveness of Network Pruning

Table 7 shows the ablation study of Lite-YOLOv5 with and without network pruning. In Table 7, "✘" means Lite-YOLOv5 without network pruning (while keeping the other five modules), while "✔" means Lite-YOLOv5 with network pruning (i.e., our proposed detector). From Table 7, one can find that network pruning can realize a lighter architecture with 4.44 G FLOPs and a 2.38 M model volume (~68.6% decrease of FLOPs and ~81.5% decline of model volume when compared with Lite-YOLOv5 without network pruning). Thus, network pruning can achieve a huge compression of the model with a slight accuracy loss, which confirms its superior performance in lightweight network design. In addition, we also conducted another experiment to explore the effect of channel-wise pruning.

**Table 7.** The ablation study of Lite-YOLOv5 with and without network pruning.


### 5.2.2. Experiment 2: Effect of Channel-Wise Pruning

During the channel pruning procedure of network pruning, we conducted several experiments under different pruning ratios *Pr*. In general, the larger the pruning ratio is, the smaller the model volume of the network is while the poorer the model performance is. Thus, it is of grea<sup>t</sup> importance to trade off the pruning ratio and the model performance. From Figure 14, we can see the effect of choosing different pruning ratios from Lite-YOLOv5 trained on LS-SSDD-v1.0 with λ = 10−3. When *Pr* goes beyond 0.7, the F1 of the model seriously deteriorates. Thus, in our implementation, Lite-YOLOv5 is channel-wise pruned with a *Pr* equal to 0.7 to trade off the model performance and model complexity.

**Figure 14.** The effect of choosing different pruning ratios from Lite-YOLOv5 trained on LS-SSDD-v1.0 with λ = 10−3: (**a**) model volume vs. F1; (**b**) FLOPs vs. F1.

### *5.3. Ablation Study on the HPBC Module*

Table 8 shows the ablation study of Lite-YOLOv5 with and without the HPBC module. In Table 8, "✘" means Lite-YOLOv5 without the HPBC module (while keeping the other five modules), while "✔" means Lite-YOLOv5 with the HPBC module (i.e., our proposed detector). From Table 8, one can find that the HBC module can make a ~0.2% improvement with AP and F1. Note that the ~0.7% improvement in precision (i.e., the decrease of false alarms) reveals the reason of the improvement of overall detection performance (i.e., the HPBC module can effectively exclude pure background ocean images, so some false alarms in them are avoided). Furthermore, the HBC module can obtain the real-time detection performance with only a 37.51 s running time for one large-scale image (~10.4 s decrease of running time compared with Lite-YOLOv5 without the HPBC module).

**Table 8.** The ablation study of Lite-YOLOv5 with and without HPBC module.


All of the above reveal that the HPBC module can effectively classify pure background ocean images; thus, it can (1) suppress some false alarms, and therefore the overall accuracy indices are increased and (2) decrease the detection burden of the detector, and therefore real-time detection performance is guaranteed. Significantly, one may find more powerful techniques to further classify the pure background samples, but HPBC might be one of the most direct approaches without complicated steps and obscure theories.

We performed another experiment to study the impact of the abscissa filter threshold *ε*a. The experimental results are shown in Table 9. It can be concluded that when εa is set higher, more pure background ocean images will be excluded (i.e., fewer images remain) and detection performance will be improved. However, in the actual scene, we focus on not excluding the ship images by mistake. Thus, it is of grea<sup>t</sup> importance to optimize εa on the basis of guaranteeing the original number of ship images. In Table 9, εa being set to 128 is the optimal choice for the balance of the number of ship images and detection accuracy. Thus, the final εa is set to 128 in Lite-YOLOv5.


**Table 9.** The ablation study of Lite-YOLOv5 with different abscissa filter thresholds. #Images: number of test set images; #Ships: number of ships in test set images.

### *5.4. Ablation Study on the SDC Module*

Table 10 shows the ablation study of Lite-YOLOv5 with and without the SDC module. In Table 10, "✘" means Lite-YOLOv5 without the SDC module (while keeping the other five modules), while "✔" means Lite-YOLOv5 with the SDC module (i.e., our proposed detector). From Table 10, one can find that the SDC module can make an overall detection performance improvement with a ~ 0.9% F1 improvement, which confirms its effectiveness. This is because the SDC module can utilize the SAR ship shape distance (i.e., the distribution of length, width, and aspect ratio) to generate a more appropriate prior anchor. Finally, Lite-YOLOv5 can extract SAR ship information more effectively. In addition, the SDC module brings hardly any model complexity increase, which confirms its superior cost-effectiveness in detection accuracy compensation.


**Table 10.** The ablation study of Lite-YOLOv5 with and without SDC module.

### *5.5. Ablation Study on the CSA Module*

Table 11 shows the ablation study of Lite-YOLOv5 with and without the CSA module. In Table 11, " ✘" means Lite-YOLOv5 without the CSA module (while keeping the other five modules), while " ✔" means Lite-YOLOv5 with the CSA module (i.e., our proposed detector). From Table 11, one can find that the CSA module can make an overall detection performance improvement with a ~2.6% AP and ~1.9% F1 improvement, which confirms its effectiveness. This is because the CSA module can extract both rich spatial and rich semantic information. Finally, Lite-YOLOv5 can improve the ship detection performance. In addition, the CSA module only brings a slight model complexity increase, which confirms its superior cost-effectiveness in detection accuracy compensation.

**Table 11.** The ablation study of Lite-YOLOv5 with and without CSA module.


### *5.6. Ablation Study on the H-SPP Module*

Table 12 shows the ablation study of Lite-YOLOv5 with and without the H-SPP module. In Table 12, " ✘" means Lite-YOLOv5 without the H-SPP module (while keeping the other five modules), while " ✔" means Lite-YOLOv5 with the H-SPP module (i.e., our proposed detector). From Table 12, one can find that the H-SPP module can make an overall detection performance improvement with a ~0.8% F1 improvement, which confirms its effectiveness. This is because the H-SPP module can aggregate the feature maps of abundant receptive fields and obtain different degrees of context information. Finally, Lite-YOLOv5 can effectively improve the network's capacity to capture both local and global information of SAR images. In addition, the H-SPP module only brings a slight model complexity increase, which confirms its superior cost-effectiveness in detection accuracy compensation.


**Table 12.** The ablation study of Lite-YOLOv5 with and without H-SPP module.
