Scene Text Erasing綜述和自己的思考 - 知乎最近在看關于Scene Text Erasing的論文，先在這挖一個坑，論文細節以后慢慢整理。 1 介紹場景文字消除（Scene Text Erasing），顧名思義為消除自然場景的文字，輸入是一張完整的帶有文字的任意場景的圖片，輸出是…https://zhuanlan.zhihu.com/p/441303427

????????從文字擦除的角度上，一般輸入有原圖，去除了文字的gt圖和包含了文字位置的mask圖，至于為什么有些帶有原始文字位置mask作為額外的監督信息是不是會更好幫助網絡定位文字，作為輔助信息，本文也為了這個維度信息。思路上有兩種方案，一種是端到端網絡，將文字定位和文字擦除當成一個整體來做，第二種是分兩步來做，將文字定位和文字擦除當成上下游任務，先解決定位文字，再將得到的文字位置信息作為先驗知識和圖片一起輸入到圖像修復中，這種想法非常普遍，ocr的文字檢測和識別，去水印等都是這個思路。SynthText，SCUT-EnsText是gt和mask的，EnsNetset是只有原圖和gt，沒有mask。其實文字擦除本質還是生成任務，現如今，這個思路上主要是圖像修復，包括超分之類的，不一定要通過生成式的GAN去做，擴散也可以做，前期的思路，通過ae或者encoder-decoder架構，用l1損失就可以很好的建模，只不過伴隨著gan的發展可控，gan思路就多起來。

1.Introduction

自監督在文字擦除或者圖像修復或者去水印上的應用是廣泛的，以Synthesis為例，輸入原圖之后，在原圖上添加水印或者其他的東西，以這張圖為輸入，原圖為gt，通過數據合成，來達到去除的目的，這中做法是非常普遍的，作者說不夠好，提出了下面的對比：

上面這張圖的尚不就是synthText的方式，從均勻分布中采樣，一般就是random.randit的這種方式，而本文提出的STE在style分布上采樣更貼合原始數據，效果更好，本質上講，生成數據和原始數據是同分布的，模型擬合這種分布學到的是更本質的知識。

? ? ? ? 提出了self-supervised Text Erasing(STE) framework。包括了兩部分，image synthesis和text erasing。在圖像合成中，除了之前的生成方法，還利用了mser提取區域近似真實圖，考慮到合成文本和真實文本之間文本風格差距，構建了policy network，reward是根據所選風格的現實性和難度來計算。在擦除模塊，是一個從粗到細的生成模型來擦除文本并用適當的紋理填充缺失的像素，并提出了對比學習的一個三元組損失。

2.self-supervised text erasing

????????這篇論文的核心在看懂這張圖，首先STE是一個自監督的框架，它的輸入只有一張圖，沒有gt，整體是一個GAN的架構，輸入I，生成器架構包括了合成函數F和一個Text Erasing Module，在經過了合成函數F之后產生了合成圖和text mask，此處的text mask是合成時添加到I上的，即Isyn，此圖輸入Text Erasing Module，輸出的是Ipred，Ipred在形式上應當與I相同，事實上輸出的Ipred即為fake_img，和real_img構成了判別器的輸入。Text Erasing Module是一個文字擦除模塊EraseNet，這個是之前就有的模型，如果有原圖及其對應的gt圖，取消了文字的圖，可以直接用來訓練，STE取了巧，把它包進了GAN這個框架里，用它做生成器，輸入合成圖，輸出是沒有文字的圖，這個過程有一個coarse model和一個refine model組成，這里是一個生成架構，但是通過什么信號來監督呢？refine model輸出圖和F添加上的text mask做乘積得到輸出圖fake_img,fake_img是包括原始的text的，refine model還輸出了一個預測的text mask，這個預測的text mask將來會在判別器里起作用，就是說整個GAN是監督原始圖生成的真不真，通過F函數是添加了text上去的，這個text的去除是在生成器中完成，并且生成器其實是祛除了所有的text的，我們在構造loss的時候給重新加回來了，Erasenet就是去text用，真實場景前向推理時只用這個模塊。

2.1 Overview

????????在STE中，I和Isyn是一對數據，利用合成數據Isyn來訓練生成模型G。為了將合成文本與原始文本對齊，使用策略網絡A為F選擇合適的樣式s，通過反饋進行優化，包括文字難度Rdiff和風格現實獎勵Rreal，Msyn表示合成區域的二進制掩碼，損失計算的結果Ipred，由Ir和Isyn以Msyn為條件合成。

2.2 style-aware synthesis function

customization mechanism:文本樣式被分解為多個單獨的單元，通過選擇每個單元中的操作，確定樣式參數。

replication mechanism:它旨在通過復制目標分布中的原始文本來合成樣本。

2.3 controllable synthesis module

2.3.1 search space

合成函數F提供了多種樣式，所有這些構成了策略網絡的搜索空間。

2.3.2 style optimization via reindorce.

策略網絡的目標是在大搜索空間中為每個圖像找到合適的合成樣式。

2.3.3 reward setting

style realistic reward:為了捕獲目標分布，實現了一個文本鑒別器Dtext來指導數據合成，具體來說，構建Dtext是為了預測合成圖像Isyn中文本區域，該模塊將生成器G的特征圖g(Isyn)作為輸入。

text difficult reward:

2.4 erasing module with triplet erasure loss

計算損失，只關注合成文本區域。

?損失函數：

基礎版本：沒有對比學習，風格控制和強化學習模塊，整體是一個GAN，GAN的判別器的真假樣本是輸入圖，就是帶有原始的text的圖，只不過在做判別器時融合了text_mask信息。

dataset,dataset_size=init_dataset-> CreateDataLoader-> CustomDatasetDataLoader.initialize()-> data_loader=CreateDataset-> dataset=ItemsDataset()-> dataset.initialize()-> imageFiles/infos->dataloader=torch.utils.data.DataLoader(dataset)-> dataset=data_loader.load()(dataloader)->train/valid/test-> create_model()（pix2pix/disc/gateconv/erase）-> model=EraseModel-> model.initialize()-> BaseModel.initialize-> set_param-> self.adaptive_feature_norm=AdaptiveFeatureNorm(0.1)-> netG=STRnet2(3)-> netD=Discriminator_STE(3)-> LossWithGAN_STE(vggfeatureextractor,netD)-> dis=torch.nn.L1loss-> self.optimizer_G=torch.optim.Adam(netG.parameters())-> netG.train()切換mode-> model.setup-> Visualizer-> model.update_learning_rate()-> ItemsDataset.__getitem__-> load_item()-> gt/info->gen_config=random_gen_config-> space_config是一個配置區間-> img=generate_img_with_config(gt,info,gen_config)-> generate_img(image,info,config)其實這就是synthesis function F-> raw_mask=gen_raw_mask text的mask->mask=get_mask()-> transformers_param=get_params()獲得一些數據預處理的參數-> trans=get_transform-> img=trans(img):512,768/mask=trans(mask)/gt=trans(gt)/raw_mask=trans(raw_mask)-> img/gt/mask/raw_mask=input_transform(img/gt/mask/raw_mask):3,768,512-> data=next(dataset_iter):gt/img/mask/raw_mask:2,3,768,512-> model.set_input(data)-> model.optimize_parameters()-> STRnet2.forward(img:2,3,768,512)-> x_o0:2,512,24,16 x_o1:2,64,192,128,x_o2:2,32,384,256,x_o3:2,3,768,512,fake_B:2,3,768,512,gen_mask:2,3,768,512,x_mask:2,256,48,32-> comp_B=fake_B*(1-mask)+real_A*mask:2,3,768,512-> mask_sigmoid-> comp_G=self.fake_B*(1-mask)+real_A*mask-> comp_all=fake_B*(1-mask)+real_A*mask*raw_mask+fake_B*(1-raw_mask)-> lossWithGAN_STE.forward(real_A輸入圖,mask/mask_gt：后續加上的mask,fake_B/output輸出的假圖，這個假圖是什么文字也沒有的圖，gen_mask/mm：這個是預測的加上去的text?mask圖，gt：和輸入圖一樣的，raw_mask：原始的text?mask的圖)-> D_real=discriminator(gt:2,3,768,512,masks:2,3,768,512)-> Discriminator_STE.forward()-> concat_feat:2,512,12,8->D_real:2,45-> D_fake=discriminator(output,mask)-> gan_mode:vanilla-> D_loss=D_real+D_fake->D_optimizer.zero_grad()-> D_loss.backward()->D_optimizer.step()-> G_fake=discriminator(output,mask)-> output_comp=mask*input+(1-mask)*output->holeLoss=10*l1((1-mask)*output,(1-mask)*gt)-> mask_loss=dice_loss(mm,1-mask_gt*raw_mask,mask_sigmoid)-> validAreaLoss-> mask:2,3,768,512;masks_a:2,3,192,128;masks_b:2,3,384,256;gt:2,3,768,512;imgs1:2,3,192,128;imgs2:2,3,384,256;raw_mask:2,3,768,512;raw_masks_a:2,3,192,128;raw_masks_b:2,3,384,192-> msr_loss-> extractor:VGG16FeatureExtractor-> feat_output_comp=extractor(output_comp)[2,64,384,256;2,128,192,128;2,256,96,64]-> feat_output=extractor(output)-> feat_gt=extractor(gt)-> prcLoss=0.01*l1(feat_output,feat_gt)-> styleLoss=120*l1(gram_matrix(feat_output),gram_matrix(feat_gt))-> GLoss=msrloss+holeLoss+validAreaLoss+prcLoss+styleLoss+G_fake+1*mask_loss-> optimizer_G.zero_grad()-> G_loss.backward()-> optimizer_G.step()

dataset,dataset_size=init_dataset->CreateDataLoader->CustomDatasetDataLoader.initialize()->data_loader=CreateDataset->dataset=ItemsDataset()->dataset.initialize()->imageFiles/infos->dataloader=torch.utils.data.DataLoader(dataset)->dataset=data_loader.load()(dataloader)->train/valid/test->create_model()（pix2pix/disc/gateconv/erase）->model=EraseModel->model.initialize()->BaseModel.initialize->set_param->self.adaptive_feature_norm=AdaptiveFeatureNorm(0.1)->netG=STRnet2(3)->netD=Discriminator_STE(3)->LossWithGAN_STE(vggfeatureextractor,netD)->dis=torch.nn.L1loss->self.optimizer_G=torch.optim.Adam(netG.parameters())->netG.train()切換mode->model.setup->Visualizer->model.update_learning_rate()->ItemsDataset.__getitem__->load_item()->gt/info->gen_config=random_gen_config->space_config是一個配置區間->img=generate_img_with_config(gt,info,gen_config)->generate_img(image,info,config)其實這就是synthesis function F->raw_mask=gen_raw_mask text的mask->mask=get_mask()->transformers_param=get_params()獲得一些數據預處理的參數->trans=get_transform->img=trans(img):512,768/mask=trans(mask)/gt=trans(gt)/raw_mask=trans(raw_mask)->img/gt/mask/raw_mask=input_transform(img/gt/mask/raw_mask):3,768,512->data=next(dataset_iter):gt/img/mask/raw_mask:2,3,768,512->model.set_input(data)->model.optimize_parameters()->STRnet2.forward(img:2,3,768,512)->x_o0:2,512,24,16 x_o1:2,64,192,128,x_o2:2,32,384,256,x_o3:2,3,768,512,fake_B:2,3,768,512,gen_mask:2,3,768,512,x_mask:2,256,48,32->comp_B=fake_B*(1-mask)+real_A*mask:2,3,768,512->mask_sigmoid->comp_G=self.fake_B*(1-mask)+real_A*mask->comp_all=fake_B*(1-mask)+real_A*mask*raw_mask+fake_B*(1-raw_mask)->lossWithGAN_STE.forward(real_A輸入圖,mask/mask_gt：后續加上的mask,fake_B/output輸出的假圖，這個假圖是什么文字也沒有的圖，gen_mask/mm：這個是預測的加上去的text?mask圖，gt：和輸入圖一樣的，raw_mask：原始的text?mask的圖)->D_real=discriminator(gt:2,3,768,512,masks:2,3,768,512)->Discriminator_STE.forward()->concat_feat:2,512,12,8->D_real:2,45->D_fake=discriminator(output,mask)->gan_mode:vanilla->D_loss=D_real+D_fake->D_optimizer.zero_grad()->D_loss.backward()->D_optimizer.step()->G_fake=discriminator(output,mask)->output_comp=mask*input+(1-mask)*output->holeLoss=10*l1((1-mask)*output,(1-mask)*gt)->mask_loss=dice_loss(mm,1-mask_gt*raw_mask,mask_sigmoid)->validAreaLoss->mask:2,3,768,512;masks_a:2,3,192,128;masks_b:2,3,384,256;gt:2,3,768,512;imgs1:2,3,192,128;imgs2:2,3,384,256;raw_mask:2,3,768,512;raw_masks_a:2,3,192,128;raw_masks_b:2,3,384,192->msr_loss->extractor:VGG16FeatureExtractor->feat_output_comp=extractor(output_comp)[2,64,384,256;2,128,192,128;2,256,96,64]->feat_output=extractor(output)->feat_gt=extractor(gt)->prcLoss=0.01*l1(feat_output,feat_gt)->styleLoss=120*l1(gram_matrix(feat_output),gram_matrix(feat_gt))->GLoss=msrloss+holeLoss+validAreaLoss+prcLoss+styleLoss+G_fake+1*mask_loss->optimizer_G.zero_grad()->G_loss.backward()->optimizer_G.step()

EraseModel->STRnet(2):

STRnet2((conv1): ConvWithActivation((conv2d): Conv2d(3, 32, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(conva): ConvWithActivation((conv2d): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(convb): ConvWithActivation((conv2d): Conv2d(32, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(res1): Residual((conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(batch_norm2d): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True))(res2): Residual((conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(batch_norm2d): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True))(res3): Residual((conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))(conv2): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(conv3): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2))(batch_norm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True))(res4): Residual((conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(batch_norm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True))(res5): Residual((conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))(conv2): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(conv3): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2))(batch_norm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True))(res6): Residual((conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(batch_norm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True))(res7): Residual((conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))(conv2): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(conv3): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2))(batch_norm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True))(res8): Residual((conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(batch_norm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True))(conv2): ConvWithActivation((conv2d): Conv2d(512, 512, kernel_size=(1, 1), stride=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(deconv1): DeConvWithActivation((conv2d): ConvTranspose2d(512, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), output_padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(deconv2): DeConvWithActivation((conv2d): ConvTranspose2d(512, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), output_padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(deconv3): DeConvWithActivation((conv2d): ConvTranspose2d(256, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), output_padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(deconv4): DeConvWithActivation((conv2d): ConvTranspose2d(128, 32, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), output_padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(deconv5): DeConvWithActivation((conv2d): ConvTranspose2d(64, 3, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), output_padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(lateral_connection1): Sequential((0): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1))(1): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(3): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1)))(lateral_connection2): Sequential((0): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1))(1): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(3): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1)))(lateral_connection3): Sequential((0): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1))(1): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(3): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1)))(lateral_connection4): Sequential((0): Conv2d(32, 32, kernel_size=(1, 1), stride=(1, 1))(1): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(3): Conv2d(64, 32, kernel_size=(1, 1), stride=(1, 1)))(conv_o1): Conv2d(64, 3, kernel_size=(1, 1), stride=(1, 1))(conv_o2): Conv2d(32, 3, kernel_size=(1, 1), stride=(1, 1))(mask_deconv_a): DeConvWithActivation((conv2d): ConvTranspose2d(512, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), output_padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(mask_conv_a): ConvWithActivation((conv2d): Conv2d(256, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(mask_deconv_b): DeConvWithActivation((conv2d): ConvTranspose2d(256, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), output_padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(mask_conv_b): ConvWithActivation((conv2d): Conv2d(128, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(mask_deconv_c): DeConvWithActivation((conv2d): ConvTranspose2d(128, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), output_padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(mask_conv_c): ConvWithActivation((conv2d): Conv2d(64, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(mask_deconv_d): DeConvWithActivation((conv2d): ConvTranspose2d(64, 32, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), output_padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(mask_conv_d): Conv2d(32, 3, kernel_size=(1, 1), stride=(1, 1))(coarse_conva): ConvWithActivation((conv2d): Conv2d(3, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(coarse_convb): ConvWithActivation((conv2d): Conv2d(32, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(coarse_convc): ConvWithActivation((conv2d): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(coarse_convd): ConvWithActivation((conv2d): Conv2d(64, 128, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(coarse_conve): ConvWithActivation((conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(coarse_convf): ConvWithActivation((conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(astrous_net): Sequential((0): ConvWithActivation((conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(2, 2), dilation=(2, 2))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(1): ConvWithActivation((conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(4, 4), dilation=(4, 4))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(2): ConvWithActivation((conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(8, 8), dilation=(8, 8))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(3): ConvWithActivation((conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(16, 16), dilation=(16, 16))(activation): LeakyReLU(negative_slope=0.2, inplace=True)))(coarse_convk): ConvWithActivation((conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(coarse_convl): ConvWithActivation((conv2d): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(coarse_deconva): DeConvWithActivation((conv2d): ConvTranspose2d(384, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), output_padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(coarse_convm): ConvWithActivation((conv2d): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(coarse_deconvb): DeConvWithActivation((conv2d): ConvTranspose2d(192, 32, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), output_padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(coarse_convn): Sequential((0): ConvWithActivation((conv2d): Conv2d(32, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(1): ConvWithActivation((conv2d): Conv2d(16, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))))(c1): Conv2d(32, 64, kernel_size=(1, 1), stride=(1, 1))(c2): Conv2d(64, 128, kernel_size=(1, 1), stride=(1, 1)) )

EraseModel:Discriminator_STE

Discriminator_STE((globalDis): Sequential((0): ConvWithActivation((conv2d): Conv2d(3, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(1): ConvWithActivation((conv2d): Conv2d(64, 128, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(2): ConvWithActivation((conv2d): Conv2d(128, 256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(3): ConvWithActivation((conv2d): Conv2d(256, 256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(4): ConvWithActivation((conv2d): Conv2d(256, 256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(5): ConvWithActivation((conv2d): Conv2d(256, 256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True)))(localDis): Sequential((0): ConvWithActivation((conv2d): Conv2d(3, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(1): ConvWithActivation((conv2d): Conv2d(64, 128, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(2): ConvWithActivation((conv2d): Conv2d(128, 256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(3): ConvWithActivation((conv2d): Conv2d(256, 256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(4): ConvWithActivation((conv2d): Conv2d(256, 256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True))(5): ConvWithActivation((conv2d): Conv2d(256, 256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))(activation): LeakyReLU(negative_slope=0.2, inplace=True)))(fusion): Sequential((0): Conv2d(512, 1, kernel_size=(4, 4), stride=(1, 1))) )

總結

以上是生活随笔為你收集整理的self-supervised text erasing with controllable image synthesis的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。