DL之DilatedConvolutions：Dilated Convolutions(膨脹卷積/擴張卷積)算法的簡介(論文介紹)、架構(gòu)詳解、案例應(yīng)用等配圖集合之詳細(xì)攻略

?

?

?

?

目錄

Dilated Convolutions算法的簡介(論文介紹)

Dilated Convolutions算法的架構(gòu)詳解

1、膨脹卷積的應(yīng)用——語義分割網(wǎng)絡(luò)中引入膨脹卷積

2、膨脹卷積的優(yōu)點

3、卷積、反卷積與膨脹卷積

Dilated Convolutions算法的案例應(yīng)用

---

?

?

?

?

Dilated Convolutions算法的簡介(論文介紹)

ABSTRACT ?
? ? ? State-of-the-art models for semantic segmentation are based on adaptations of ?convolutional networks that had originally been designed for image classification. ?However, dense prediction problems such as semantic segmentation are ?structurally different from image classification. In this work, we develop a new ?convolutional network module that is specifically designed for dense prediction. ?The presented module uses dilated convolutions to systematically aggregate multiscale ?contextual information without losing resolution. The architecture is based ?on the fact that dilated convolutions support exponential expansion of the receptive ?field without loss of resolution or coverage. We show that the presented context ?module increases the accuracy of state-of-the-art semantic segmentation systems. ?In addition, we examine the adaptation of image classification networks to dense ?prediction and show that simplifying the adapted network can increase accuracy.
? ? ? 最先進的語義分割模型是基于卷積網(wǎng)絡(luò)的自適應(yīng)，而卷積網(wǎng)絡(luò)最初是為圖像分類而設(shè)計的。然而，語義分割等密集預(yù)測問題在結(jié)構(gòu)上與圖像分類不同。在這項工作中，我們開發(fā)了一個新的卷積網(wǎng)絡(luò)模塊，專門為密集預(yù)測設(shè)計。所提出的模組使用擴展卷積來系統(tǒng)地聚合多尺度的上下文信息而不丟失分辨率。該架構(gòu)基于這樣一個事實，即膨脹的卷積支持接收域的指數(shù)級擴展，而不會丟失分辨率或覆蓋率。結(jié)果表明，提出的上下文模塊提高了目前最先進的語義分割系統(tǒng)的精度。此外，我們研究了圖像分類網(wǎng)絡(luò)對密集預(yù)測的適應(yīng)性，并證明簡化自適應(yīng)網(wǎng)絡(luò)可以提高精度。
CONCLUSION ?
? ? ? We have examined convolutional network architectures for dense prediction. Since the model must ?produce high-resolution output, we believe that high-resolution operation throughout the network is both feasible and desirable. Our work shows that the dilated convolution operator is particularly ?suited to dense prediction due to its ability to expand the receptive field without losing resolution ?or coverage. We have utilized dilated convolutions to design a new network structure that reliably ?increases accuracy when plugged into existing semantic segmentation systems. As part of this work, ?we have also shown that the accuracy of existing convolutional networks for semantic segmentation ?can be increased by removing vestigial components that had been developed for image classification.
? ? ? 我們研究了用于密集預(yù)測的卷積網(wǎng)絡(luò)架構(gòu)。由于模型必須產(chǎn)生高分辨率的輸出，我們認(rèn)為整個網(wǎng)絡(luò)的高分辨率操作是可行的，也是可取的。我們的工作表明，膨脹卷積算子特別適合于密集預(yù)測，因為它能夠在不損失分辨率或覆蓋率的情況下擴展接收域。我們利用擴展卷積設(shè)計了一種新的網(wǎng)絡(luò)結(jié)構(gòu)，當(dāng)插入現(xiàn)有的語義分割系統(tǒng)時，可以可靠地提高精確度。作為這項工作的一部分，我們還表明，通過去除用于圖像分類的殘留成分，可以提高現(xiàn)有卷積網(wǎng)絡(luò)用于語義分割的準(zhǔn)確性。
? ? ? We believe that the presented work is a step towards dedicated architectures for dense prediction that ?are not constrained by image classification precursors. As new sources of data become available, ?future architectures may be trained densely end-to-end, removing the need for pre-training on image ?classification datasets. This may enable architectural simplification and unification. Specifically, ?end-to-end dense training may enable a fully dense architecture akin to the presented context network ?to operate at full resolution throughout, accepting the raw image as input and producing dense ?label assignments at full resolution as output. ?
? ? ? 我們認(rèn)為，所提出的工作是朝著不受圖像分類前驅(qū)體約束的高密度預(yù)測專用體系結(jié)構(gòu)邁進的一步。隨著新數(shù)據(jù)源的出現(xiàn)，未來的體系結(jié)構(gòu)可能需要密集的端到端培訓(xùn)，從而無需對圖像分類數(shù)據(jù)集進行預(yù)培訓(xùn)。這可能使架構(gòu)簡化和統(tǒng)一成為可能。具體地說，端到端密集訓(xùn)練可能使類似于所述上下文網(wǎng)絡(luò)的完全密集的體系結(jié)構(gòu)能夠以全分辨率運行，接受原始圖像作為輸入，并以全分辨率生成密集的標(biāo)簽分配作為輸出。
? ? ? State-of-the-art systems for semantic segmentation leave significant room for future advances. Failure ?cases of our most accurate configuration are shown in Figure 4. We will release our code and ?trained models to support progress in this area.
? ? ? 最先進的語義分割系統(tǒng)為未來的發(fā)展留下了巨大的空間。圖4顯示了我們最精確配置的故障案例。我們將發(fā)布我們的代碼和經(jīng)過培訓(xùn)的模型來支持這一領(lǐng)域的進展。

?

論文
Fisher Yu, VladlenKoltun.
Multi-Scale Context Aggregation by Dilated Convolutions. ICLR, 2016
https://arxiv.org/abs/1511.07122

?

?

?

Dilated Convolutions算法的架構(gòu)詳解

更新……

1、卷積、反卷積與膨脹卷積

DL之CNN：卷積神經(jīng)網(wǎng)絡(luò)算法簡介之卷積矩陣、轉(zhuǎn)置卷積(反卷積Transpose)、膨脹卷積(擴張卷積Dilated)、帶孔卷積(atrous )之詳細(xì)攻略

?

?

Dilated Convolutions算法的案例應(yīng)用

更新……

?

?

?

?

?

?

?

?

?

?

?

?

?

?

《新程序員》：云原生和全面數(shù)字化實踐50位技術(shù)專家共同創(chuàng)作，文字、視頻、音頻交互閱讀

總結(jié)

以上是生活随笔為你收集整理的DL之DilatedConvolutions：Dilated Convolutions(膨胀卷积/扩张卷积)算法的简介(论文介绍)、架构详解、案例应用等配图集合之详细攻略的全部內(nèi)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(wǎng)站內(nèi)容還不錯，歡迎將生活随笔推薦給好友。