*4.4. Ablation Study*

#### 4.4.1. Effects of MPANet

We conduct several ablation studies on ICDAR2015 and CTW1500 datasets to verify the effectiveness of the proposed MPANet(see Table 2). Note that all the models are trained using only official training images. As shown in Table 2, MPANet obtains 1.01% and 1.21% improvement in F-measure on ICDAR2015 and CTW1500, respectively.

**Table 2.** The performance gain of MPANet. \* and † are results from ICDAR2015 and CTW1500, respectively. FPN \* and FPN † represent the results of using the FPN network model in PSE [21] on ICDAR2015 and CTW1500, respectively.


Figure 4 shows the train loss difference between modified PANet with SE block (SEMPANet) and MPANet without SE block (MPANet). It demonstrates that the loss function of SEMPANet drops faster on ICDAR2015. Figure 5 shows the loss comparison of two models with and without SE block, which proves that the loss function of MPANet model has a slightly faster convergence effect on average than the other one on CTW1500. The difference of the loss function on the two datasets is reflected in the last two rows of Table 4 and Table 5.

**Figure 4.** Ablation study of an SE block on ICDAR2015. These results are based on (ResNet 50 and SE block) and (ResNet 50 block) trained on MPANet.

**Figure 5.** Ablation study of an SE block on CTW1500. These results are based on (ResNet 50 and SE block) and (ResNet 50 block) trained on MPANet.

4.4.2. Effects of the Threshold *λ* in the Testing Phase

The hyper-parameter *λ* in the final test balances the influence between the three evaluation indexes. Table 3 compares the prediction effects of MPANet and SEMPANet with different *λ* within a short fluctuation range on the dataset ICDAR2015. We see that when SEMPANet with a *λ* of 0.89 is used, even if the performance is robust to changes in *λ*, in the average performance of the three evaluation indexes, F-measure is higher than PSENet, and Recall also performs best.


**Table 3.** The performance comparison of *λ*.
