人臉驗證（四）--CenterLoss，損失函式改進原因 人臉驗證（四）--CenterLoss
阿新 • • 發佈：2018-11-28
轉
人臉驗證（四）--CenterLoss
2017年12月10日 11:44:31 EasonApp 閱讀數：918有點糾結，實質上本文只是在深度學習框架下人臉識別損失函式的改進。但我還是把它歸類到端到端了。出於個人原因。
轉自：http://blog.csdn.net/yang_502/article/details/72792786
16年ECCV的文章《A Discriminative Feature Learning Approach for Deep Face Recognition》
code：https://github.com/ydwen/caffe-face
Motivation：
和metric learning的想法一致，希望同類樣本之間緊湊，不同類樣本之間分散。現有的CNN最常用的softmax損失函式來訓練網路，得到的深度特徵通常具有比較強的區分性，也就是比較強的類間判別力。關於softmax的類內判別力，作者在文章中給了toy example,很直觀的理解。
上面給的是mnist的最後一層特徵在二維空間的一個分佈情況，可以看到類間是可分的，但類記憶體在的差距還是比較大的。對於像人臉這種複雜分佈的資料，我們通常不僅希望資料在特徵空間不僅是類間可分，更重要的是類內緊湊。因為同一個人的類內變化很可能會大於類間的變化，只有保持類內緊湊，我們才能對那些類內大變化的樣本有一個更加魯棒的判定結果。也就是學習一種discriminative的特徵。
下圖就是我們希望達到的一種效果。
另外一種通俗解釋 ：
一句話：通過新增center loss 讓簡單的softmax 能夠訓練出更有內聚性的特徵
作者意圖，在配合softmax適用的時候，希望使學習到的特徵具有更好的泛化性和辨別能力。通過懲罰每個種類的樣本和該種類樣本中心的偏移，使得同一種類的樣本儘量聚合在一起。
相對於triplet（Google FaceNet: A Unified Embedding for Face Recognition and Clustering：2015）和contrastive(湯曉鷗 Deep LearningFace Representation by Joint Identification-Verification：2014)來說，這個目標其實相對‘清晰’， 所以網路收斂的速度甚至比僅僅用softmax更快，而且不需要像前兩者那樣構造大量的訓練對。
看一張圖：
在左圖中，我們發現一個類如果太胖，那麼出現的結果就是類內距離>類間距離。（任意紅點之間的距離應該小於紅藍之間的距離。）
左邊時softmax一般結果，右邊時centerloss結果。我們期望特徵不僅可分，而且時必須差異大。如右邊圖。
Approach：
考慮保持softmax loss的類間判別力，提出center loss，center loss就是為了約束類內緊湊的條件。相比於傳統的CNN，僅改變了原有的損失函式，易於訓練和優化網路。
softmax loss :
center loss:
其中，cyi就是每類人特徵的中心表示。這個就是一個類中心特徵，希望在一類人的所有影象特徵與類中心特徵的距離總和最小。
完整的loss function：
關於center loss 的求導：
最後的loss實現可以參考code的具體實現。
Experiments：
CNN architecture
Weights in three local convolution layers are locally shared in the regions of 4 × 4, 2 × 2 and 1 × 1 respectively.
Implementation Details
Preprocessing.
使用【40】的關鍵點檢測，進行人臉對齊與檢測。人臉影象裁剪成112*96大小，畫素歸一化操作是通過減127.5再除128（？？沒搞懂）
Training data.
文章中使用的訓練資料比較多，包括有CASIA-WebFace [39], CACD2000 [4], Celebrity+ [22]。但是作者他們做了影象清理，保證和測試集的人身份不重複，最後保留17,189個人的0.7M的影象。後面的訓練階段，只用了0.49M的資料，並且對影象做了水平翻轉的資料增強。
Detailed settings in CNNs.
文章比較了3中模型：we respectively train three kind of models under the supervision of softmax loss (model A), softmax loss and contrastive loss (model B), softmax loss and center loss (model C)。就是網路使用的損失函式不同。
batchsize=256，在Titan X的兩塊GPU上做訓練。A和C的初始lr=0.1,分別在16K和24K之後縮小至0.01.文章提到Ａ在２８Ｋ次收斂大致花１４h,model B的收斂比較慢，初始lr=.01,咋２４Ｋ和３６Ｋ迭代的時候縮小學習率。Ｂ總共迭代了４２Ｋ次，花２２h.沒說到Ｃ的訓練時間。
Detailed settings in testing.
測試階段，將測試影象通過ＣＮＮ提取深度特徵，即最後一個ＦＣ層的輸出，將原圖以及原圖的水平翻轉圖的特徵圖拼接表示成共同的表示，經過ＰＣＡ降維後計算ＣＯＳ距離。對於識別或者驗證任務，可以是簡單通過閾值劃分或者是近鄰判定。
Experiments on the LFW and YTF datasets
使用０．７Ｍ的訓練資料，在ＬＦＷ和ＹＴＦ的pair對上做了人臉驗證。
Experiments on the dataset of MegaFace Challenge
使用ＭｅｇａＦａｃｅ這個百萬級別的資料庫（干擾影象很多）。這個對於識別任務增大了難度。
資料庫具體包括：
MegaFace datasets include gallery set and probe set. The gallery set consists of more than 1 million images from 690K different individuals, as a subset of Flickr photos [35] from Yahoo. The probe set using in this challenge are two existing databases: Facescrub [24] and FGNet [1]. Facescrub dataset is publicly available dataset, containing 100K photos of 530 unique individuals (55,742 images of males and 52,076 images of females). The possible bias can be reduced by sufficient samples in each identity. FGNet dataset is a face aging dataset, with 1002 images from 82 identities. Each identity has multiple face images at different ages (ranging from 0 to 69).
Ｒｅｆｅｒｅｎｃｅ
【４０】 ＭＴＣＮＮ
Zhang, K., Zhang, Z., Li, Z., Qiao, Y.: Joint face detection and alignment using multi-task cascaded convolutional networks. arXiv preprint arXiv:1604.02878(2016)
<script>
(function(){
function setArticleH(btnReadmore,posi){
var winH = $(window).height();
var articleBox = $("div.article_content");
var artH = articleBox.height();
if(artH > winH*posi){
articleBox.css({
'height':winH*posi+'px',
'overflow':'hidden'
})
btnReadmore.click(function(){
articleBox.removeAttr("style");
$(this).parent().remove();
})
}else{
btnReadmore.parent().remove();
}
}
var btnReadmore = $("#btn-readmore");
if(btnReadmore.length>0){
if(currentUserName){
setArticleH(btnReadmore,3);
}else{
setArticleH(btnReadmore,1.2);
}
}
})()
</script>
</article>
<div class="comment-edit-box d-flex">
<a id="commentsedit"></a>
<div class="user-img">
<a href="javascript:void(0);" target="_blank">
<img class="show_loginbox" src="//g.csdnimg.cn/static/user-img/anonymous-User-img.png">
</a>
</div>
<form id="commentform">
<input type="hidden" id="comment_replyId">
<textarea class="comment-content" name="comment_content" id="comment_content" placeholder="想對作者說點什麼"></textarea>
<div class="opt-box"> <!-- d-flex -->
<div id="ubbtools" class="add_code">
<a href="#insertcode" code="code" target="_self"><i class="icon iconfont icon-daima"></i></a>
</div>
<input type="hidden" id="comment_replyId" name="comment_replyId">
<input type="hidden" id="comment_userId" name="comment_userId" value="">
<input type="hidden" id="commentId" name="commentId" value="">
<div style="display: none;" class="csdn-tracking-statistics tracking-click" data-mod="popu_384"><a href="#" target="_blank" class="comment_area_btn">發表評論</a></div>
<div class="dropdown" id="myDrap">
<a class="dropdown-face d-flex align-items-center" data-toggle="dropdown" role="button" aria-haspopup="true" aria-expanded="false">
<div class="txt-selected text-truncate">新增程式碼片</div>
<svg class="icon d-block" aria-hidden="true">
<use xlink:href="#csdnc-triangledown"></use>
</svg>
</a>
<ul class="dropdown-menu" id="commentCode" aria-labelledby="drop4">
<li><a data-code="html">HTML/XML</a></li>
<li><a data-code="objc">objective-c</a></li>
<li><a data-code="ruby">Ruby</a></li>
<li><a data-code="php">PHP</a></li>
<li><a data-code="csharp">C</a></li>
<li><a data-code="cpp">C++</a></li>
<li><a data-code="javascript">JavaScript</a></li>
<li><a data-code="python">Python</a></li>
<li><a data-code="java">Java</a></li>
<li><a data-code="css">CSS</a></li>
<li><a data-code="sql">SQL</a></li>
<li><a data-code="plain">其它</a></li>
</ul>
</div>
<div class="right-box">
<span id="tip_comment" class="tip">還能輸入<em>1000</em>個字元</span>
<input type="submit" class="btn btn-sm btn-red btn-comment" value="發表評論">
</div>
</div>
</form>
</div>
<div class="comment-list-container">
<a id="comments"></a>
<div class="comment-list-box">
</div>
<div id="commentPage" class="pagination-box d-none"></div>
</div>
用基於center loss的人臉識別模型對LFW人臉資料集進行評測（c++） - 花澤
05-13 6681
接上一篇博文，這篇博文主要是進行人臉識別中的第③和第四個步驟：特徵提取以及相似度計算。 center loss是2016的一篇ECCV論文中提出來的，A Discrimina... 來自： 花澤
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/oppo62258801/article/details/78034014,BlogCommendClickRateRank_1,index_1"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/oppo62258801/article/details/78034014,BlogCommendClickRateRank_1,index_1"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/oppo62258801/article/details/78034014" target="_blank" title="損失函式：Hinge Loss（max margin） - oppo62258801的部落格">
<h4 class="text-truncate oneline" style="width: 610px;">
損失函式：Hinge Loss（max margin） - oppo62258801的部落格 </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/D/7/3/3_oppo62258801.jpg" alt="oppo62258801" class="avatar-pic">
<span class="namebox">
<span class="name">oppo62258801</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">09-19</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
1524</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/oppo62258801/article/details/78034014" target="_blank" title="損失函式：Hinge Loss（max margin） - oppo62258801的部落格">
<span class="desc oneline">目錄(?)[-]
Hinge Loss簡介變種
Hinge Loss簡介
Hinge Loss是一種目標函式（或者說損失函式）的名稱，有的時候又叫做max-margin …
來自： oppo62258801的部落格
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/wfei101/article/details/79773547,BlogCommendClickRateRank_2,index_2"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/wfei101/article/details/79773547,BlogCommendClickRateRank_2,index_2"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/wfei101/article/details/79773547" target="_blank" title="人臉識別：Loss函式總結 - BigCowPeking">
<h4 class="text-truncate oneline" style="width: 610px;">
<em>人臉</em>識別：Loss函式總結 - BigCowPeking </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/3/8/D/3_wfei101.jpg" alt="wfei101" class="avatar-pic">
<span class="namebox">
<span class="name">wfei101</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">03-31</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
1697</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/wfei101/article/details/79773547" target="_blank" title="人臉識別：Loss函式總結 - BigCowPeking">
<span class="desc oneline">基於深度學習的人臉識別發展，從deepid開始，到今年（或者說去年），已經基本趨於成熟。凡是基於識別的，總是離不開三個東西：資料，網路，以及loss。資料方面，目前的公開資料集中有主打類別數的MS_c...</span>
</a>
<span class="blog_title_box oneline"><a target="_blank" href="https://blog.csdn.net/wfei101">來自： <span class="blog_title"> BigCowPeking</span></a></span>
</p>
</div>
</div>
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/tinyzhao/article/details/53236191,BlogCommendClickRateRank_3,index_3"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/tinyzhao/article/details/53236191,BlogCommendClickRateRank_3,index_3"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/tinyzhao/article/details/53236191" target="_blank" title="人臉檢測：MTCNN - flyingzhao">
<h4 class="text-truncate oneline" style="width: 608px;">
<em>人臉</em>檢測：MTCNN - flyingzhao </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/7/3/0/3_tinyzhao.jpg" alt="tinyzhao" class="avatar-pic">
<span class="namebox">
<span class="name">tinyzhao</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">11-19</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
3.2萬</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/tinyzhao/article/details/53236191" target="_blank" title="人臉檢測：MTCNN - flyingzhao">
<span class="desc oneline">《Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks》論文解讀。本文來自於中國科學院...</span>
</a>
<span class="blog_title_box oneline"><a target="_blank" href="https://blog.csdn.net/tinyzhao">來自： <span class="blog_title"> flyingzhao</span></a></span>
</p>
</div>
</div>
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/angus_17/article/details/78894180,BlogCommendClickRateRank_4,index_4"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/angus_17/article/details/78894180,BlogCommendClickRateRank_4,index_4"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/angus_17/article/details/78894180" target="_blank" title="人臉識別領域中閾值的計算 - 漂泊小柒的專欄">
<h4 class="text-truncate oneline" style="width: 610px;">
<em>人臉</em>識別領域中閾值的計算 - 漂泊小柒的專欄 </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/D/8/C/3_angus_17.jpg" alt="angus_17" class="avatar-pic">
<span class="namebox">
<span class="name">angus_17</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">12-25</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
1033</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/angus_17/article/details/78894180" target="_blank" title="人臉識別領域中閾值的計算 - 漂泊小柒的專欄">
<span class="desc oneline">舉一個實際案例，機場安檢前臺，用的1:1核身比對，正樣本是正常安檢使用者，負樣本是持他人護照的客戶（攻擊！）
會這麼思考問題：
a.這個場景下，正樣本遠遠多於負樣本（攻擊極少），比如十萬比一；
b…
來自： 漂泊小柒的專欄
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/wfei101/article/details/79252067,BlogCommendClickRateRank_5,index_5"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/wfei101/article/details/79252067,BlogCommendClickRateRank_5,index_5"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/wfei101/article/details/79252067" target="_blank" title="人臉識別：損失函式之softmax loss和cross entropy Loss - BigCowPeking">
<h4 class="text-truncate oneline" style="width: 610px;">
<em>人臉</em>識別：損失函式之softmax loss和cross entropy Loss - BigCowPeking </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/3/8/D/3_wfei101.jpg" alt="wfei101" class="avatar-pic">
<span class="namebox">
<span class="name">wfei101</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">02-04</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
1404</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/wfei101/article/details/79252067" target="_blank" title="人臉識別：損失函式之softmax loss和cross entropy Loss - BigCowPeking">
<span class="desc oneline">轉載blog：http://blog.csdn.net/u014380165/article/details/77284921
我們知道卷積神經網路（CNN）在影象領域的應用已經非常廣泛了，一般…
來自： BigCowPeking
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/qq_14845119/article/details/53308996,BlogCommendClickRateRank_6,index_6"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/qq_14845119/article/details/53308996,BlogCommendClickRateRank_6,index_6"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/qq_14845119/article/details/53308996" target="_blank" title="人臉識別之caffe-face - 年輕即出發，">
<h4 class="text-truncate oneline" style="width: 608px;">
<em>人臉</em>識別之caffe-face - 年輕即出發， </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/5/B/9/3_qq_14845119.jpg" alt="qq_14845119" class="avatar-pic">
<span class="namebox">
<span class="name">qq_14845119</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">11-23</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
1.9萬</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/qq_14845119/article/details/53308996" target="_blank" title="人臉識別之caffe-face - 年輕即出發，">
<span class="desc oneline">該論文為2016年的一篇ECCV，ADiscriminative Feature Learning Approach for Deep Face Recognition。還是深圳先進院喬宇老師組的一篇...</span>
</a>
<span class="blog_title_box oneline"><a target="_blank" href="https://blog.csdn.net/qq_14845119">來自： <span class="blog_title"> 年輕即出發，</span></a></span>
</p>
</div>
</div>
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/dreamer_on_air/article/details/78521071,BlogCommendClickRateRank_7,index_7"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/dreamer_on_air/article/details/78521071,BlogCommendClickRateRank_7,index_7"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/dreamer_on_air/article/details/78521071" target="_blank" title="人臉識別：coco loss-Rethinking Feature Discrimination and Polymerization for Large-scale Recognition - 迷途中一個小書童">
<h4 class="text-truncate oneline" style="width: 610px;">
<em>人臉</em>識別：coco loss-Rethinking Feature Discrimination and Polymerization for Large-scale Recognition - 迷途中一個小書童 </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/8/A/D/3_dreamer_on_air.jpg" alt="dreamer_on_air" class="avatar-pic">
<span class="namebox">
<span class="name">dreamer_on_air</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">11-13</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
1443</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/dreamer_on_air/article/details/78521071" target="_blank" title="人臉識別：coco loss-Rethinking Feature Discrimination and Polymerization for Large-scale Recognition - 迷途中一個小書童">
<span class="desc oneline">nips的一篇做分類和識別的工作，其中在人臉識別任務上也做了實驗，Rethinking Feature Discrimination and Polymerization for Large-scal...</span>
</a>
<span class="blog_title_box oneline"><a target="_blank" href="https://blog.csdn.net/dreamer_on_air">來自： <span class="blog_title"> 迷途中一個小書童</span></a></span>
</p>
</div>
</div>
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/wfei101/article/details/79773493,BlogCommendClickRateRank_8,index_8"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/wfei101/article/details/79773493,BlogCommendClickRateRank_8,index_8"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/wfei101/article/details/79773493" target="_blank" title="目標檢測之Loss：Center Loss梯度更新 - BigCowPeking">
<h4 class="text-truncate oneline" style="width: 618px;">
目標檢測之Loss：Center Loss梯度更新 - BigCowPeking </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/3/8/D/3_wfei101.jpg" alt="wfei101" class="avatar-pic">
<span class="namebox">
<span class="name">wfei101</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">03-31</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
442</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/wfei101/article/details/79773493" target="_blank" title="目標檢測之Loss：Center Loss梯度更新 - BigCowPeking">
<span class="desc oneline">轉載：https://blog.csdn.net/u014380165/article/details/76946339最近幾年網路效果的提升除了改變網路結構外，還有一群人在研究損失層的改進，這篇博文...</span>
</a>
<span class="blog_title_box oneline"><a target="_blank" href="https://blog.csdn.net/wfei101">來自： <span class="blog_title"> BigCowPeking</span></a></span>
</p>
</div>
</div>
<div class="recommend-item-box type_hot_word">
<div class="content clearfix oneline" style="width: 702px;">
<h5 class="float-left">文章熱詞</h5>
<div class="word float-left">
<span>
<a href="https://edu.csdn.net/courses/o5329_s5330_k " target="_blank">
機器學習 </a></span>
<span>
<a href="https://edu.csdn.net/courses/o5329_s5330_k " target="_blank">
機器學習課程 </a></span>
<span>
<a href="https://edu.csdn.net/courses/o5329_s5330_k " target="_blank">
機器學習教程 </a></span>
<span>
<a href="https://edu.csdn.net/combos/o5329_s5331_l0_t " target="_blank">
深度學習視訊教程 </a></span>
<span>
<a href="https://edu.csdn.net/combos/o5329_s5331_l0_t " target="_blank">
深度學習學習 </a></span>
</div>
</div>
<div class="content clearfix oneline" style="width: 702px;">
<h5 class="float-left">相關熱詞</h5>
<div class="float-left">
<span>
<a href="https://blog.csdn.net/qq_17278169/article/details/79510209" target="_blank">
人臉</a>
</span>
<span>
<a href="https://blog.csdn.net/Trent1985/article/details/79760854" target="_blank">
人臉檄文人臉</a>
</span>
<span>
<a href="https://blog.csdn.net/c2a2o2/article/details/72648876" target="_blank">
按照人臉生成人臉</a>
</span>
<span>
<a href="https://blog.csdn.net/wfei101/article/details/80489119" target="_blank">
人臉模型檢測人臉</a>
</span>
<span>
<a href="https://blog.csdn.net/as_far_as/article/details/81391867" target="_blank">
百度人臉V3人臉</a>
</span>
</div>
</div>
</div>
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/Fire_Light_/article/details/79598497,BlogCommendClickRateRank_9,index_9"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/Fire_Light_/article/details/79598497,BlogCommendClickRateRank_9,index_9"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/Fire_Light_/article/details/79598497" target="_blank" title="人臉識別系列（十二）：Center Loss - Fire_Light_的部落格">
<h4 class="text-truncate oneline" style="width: 610px;">
<em>人臉</em>識別系列（十二）：Center Loss - Fire_Light_的部落格 </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/8/2/5/3_fire_light_.jpg" alt="Fire_Light_" class="avatar-pic">
<span class="namebox">
<span class="name">Fire_Light_</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">03-18</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
1378</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/Fire_Light_/article/details/79598497" target="_blank" title="人臉識別系列（十二）：Center Loss - Fire_Light_的部落格">
<span class="desc oneline">論文連結：A Discriminative Feature Learning Approach for Deep Face Recognition（需翻牆）
程式碼連結:https://github…
來自： Fire_Light_的部落格
<div class="recommend-item-box blog-expert-recommend-box" style="display: block;">
<div class="d-flex">
<div class="blog-expert-recommend">
<div class="blog-expert">
<div class="blog-expert-flexbox" data-track-view="{"mod":"popu_709","con":"https://blog.csdn.net/app_12062011/article/details/78763892"}"><div class="blog-expert-item"><div class="blog-expert-info-box"><div class="blog-expert-img-box" data-track-click="{"mod":"popu_709","con":"https://blog.csdn.net/app_12062011/article/details/78763892"}"><a href="https://blog.csdn.net/App_12062011" target="_blank"><img src="https://avatar.csdn.net/9/0/8/3_app_12062011.jpg" alt="EasonApp"><svg class="icon" aria-hidden="true"><use xlink:href="#csdnc-blogexpert"></use></svg></a></div><div class="info"><span data-track-click="{"mod":"popu_709","con":"https://blog.csdn.net/app_12062011/article/details/78763892"}"><a href="https://blog.csdn.net/App_12062011" target="_blank"><h5 class="oneline">EasonApp</h5></a></span> <p><span data-track-click="{"mod":"popu_710","con":"https://blog.csdn.net/app_12062011/article/details/78763892"}"><span class="blog-expert-button-follow btn-red-follow" data-name="App_12062011" data-nick="EasonApp">關注</span></span><span class="article-num"> 590篇文章</span></p></div></div></div><div class="blog-expert-item"><div class="blog-expert-info-box"><div class="blog-expert-img-box" data-track-click="{"mod":"popu_709","con":"https://blog.csdn.net/app_12062011/article/details/78763892"}"><a href="https://blog.csdn.net/zoroduyu" target="_blank"><img src="https://avatar.csdn.net/3/1/8/3_zoroduyu.jpg" alt="Forward_duyu"></a></div><div class="info"><span data-track-click="{"mod":"popu_709","con":"https://blog.csdn.net/app_12062011/article/details/78763892"}"><a href="https://blog.csdn.net/zoroduyu" target="_blank"><h5 class="oneline">Forward_duyu</h5></a></span> <p><span data-track-click="{"mod":"popu_710","con":"https://blog.csdn.net/app_12062011/article/details/78763892"}"><span class="blog-expert-button-follow btn-red-follow" data-name="zoroduyu" data-nick="Forward_duyu">關注</span></span><span class="article-num"> 15篇文章</span></p></div></div></div><div class="blog-expert-item"><div class="blog-expert-info-box"><div class="blog-expert-img-box" data-track-click="{"mod":"popu_709","con":"https://blog.csdn.net/app_12062011/article/details/78763892"}"><a href="https://blog.csdn.net/Muzythoof" target="_blank"><img src="https://avatar.csdn.net/1/E/9/3_muzythoof.jpg" alt="Muzythoof"></a></div><div class="info"><span data-track-click="{"mod":"popu_709","con":"https://blog.csdn.net/app_12062011/article/details/78763892"}"><a href="https://blog.csdn.net/Muzythoof" target="_blank"><h5 class="oneline">Muzythoof</h5></a></span> <p><span data-track-click="{"mod":"popu_710","con":"https://blog.csdn.net/app_12062011/article/details/78763892"}"><span class="blog-expert-button-follow btn-red-follow" data-name="Muzythoof" data-nick="Muzythoof">關注</span></span><span class="article-num"> 4篇文章</span></p></div></div></div></div>
</div>
</div>
<div class="blog-expert-load-new" data-index="3" style="display: block;">
<svg version="1.1" viewBox="0 0 200 200" style="enable-background:new 0 0 200 200;" xml:space="preserve">
<path d="M142.5,57.5c-1.3-1.2-2.1-3.3-2.1-5.4c0-4.2,3.3-7.5,7.5-7.5h19.6c4.2,0,7.5,3.3,7.5,7.5s-3.3,7.5-7.5,7.5h-4.2
C170.8,71.2,175,85,175,100c0,41.7-33.8,75-75,75c-4.2,0-7.5-3.3-7.5-7.5c0-4.2,3.3-7.5,7.5-7.5c33.3,0,60-27.1,60-60
C160,83.7,153.8,68.7,142.5,57.5L142.5,57.5z M57.5,142.5c1.2,1.2,2.1,3.3,2.1,5.4c0,4.2-3.3,7.5-7.5,7.5H32.5
c-4.2,0-7.5-3.3-7.5-7.5c0-4.2,3.3-7.5,7.5-7.5h4.2C29.2,128.8,25,115,25,100c0-41.7,33.7-75,75-75c4.2,0,7.5,3.3,7.5,7.5
S104.2,40,100,40c-33.3,0-60,27.1-60,60C40,116.2,46.2,131.2,57.5,142.5z"></path>
</svg>
<span>換一批</span>
</div>
</div>
</div>
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/gdmmzmj/article/details/76100051,BlogCommendClickRateRank_10,index_10"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/gdmmzmj/article/details/76100051,BlogCommendClickRateRank_10,index_10"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/gdmmzmj/article/details/76100051" target="_blank" title="centerloss 理解 - 專注於機器學習，深度學習，人臉識別領域。">
<h4 class="text-truncate oneline" style="width: 610px;">
<em>centerloss</em> 理解 - 專注於機器學習，深度學習，<em>人臉</em>識別領域。 </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/6/3/B/3_gdmmzmj.jpg" alt="gdmmzmj" class="avatar-pic">
<span class="namebox">
<span class="name">gdmmzmj</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">07-26</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
4536</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/gdmmzmj/article/details/76100051" target="_blank" title="centerloss 理解 - 專注於機器學習，深度學習，人臉識別領域。">
<span class="desc oneline">今天看了一下centerloss 原始碼 ，覺得還是挺值得借鑑的，所以分享一下。
centerloss主要在caffe的基礎上加了三項。一. caffe.proto1.加proto 格式 的 ID…
來自： 專注於機器學習，深度學習，人臉識別領域。
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/linolzhang/article/details/71075756,BlogCommendClickRateRank_11,index_11"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/linolzhang/article/details/71075756,BlogCommendClickRateRank_11,index_11"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/linolzhang/article/details/71075756" target="_blank" title="基於深度學習的Person Re-ID（綜述） - 跟隨技術的腳步-linolzhang的專欄">
<h4 class="text-truncate oneline" style="width: 608px;">
基於深度學習的Person Re-ID（綜述） - 跟隨技術的腳步-linolzhang的專欄 </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/F/A/6/3_linolzhang.jpg" alt="linolzhang" class="avatar-pic">
<span class="namebox">
<span class="name">linolzhang</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">05-01</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
1.2萬</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/linolzhang/article/details/71075756" target="_blank" title="基於深度學習的Person Re-ID（綜述） - 跟隨技術的腳步-linolzhang的專欄">
<span class="desc oneline">一. 問題的提出
Person Re-ID 全稱是 Person Re-Identification，又稱為
行人重檢測 or 行人再識別，直觀上可以通過兩種思路進行比對，一種是 通過…
來自： 跟隨技術的腳步-linolzhang的專欄
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/wfei101/article/details/79251987,BlogCommendClickRateRank_12,index_12"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/wfei101/article/details/79251987,BlogCommendClickRateRank_12,index_12"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/wfei101/article/details/79251987" target="_blank" title="人臉識別：損失函式之softmax函式 - BigCowPeking">
<h4 class="text-truncate oneline" style="width: 610px;">
<em>人臉</em>識別：損失函式之softmax函式 - BigCowPeking </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/3/8/D/3_wfei101.jpg" alt="wfei101" class="avatar-pic">
<span class="namebox">
<span class="name">wfei101</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">02-04</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
1035</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/wfei101/article/details/79251987" target="_blank" title="人臉識別：損失函式之softmax函式 - BigCowPeking">
<span class="desc oneline">Loss Function
softmax_loss的計算包含2步：
（1）計算softmax歸一化概率
（2）計算損失
這裡以batchsize=1的2分類為…
來自： BigCowPeking
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/u011747846/article/details/80602786,BlogCommendClickRateRank_13,index_13"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/u011747846/article/details/80602786,BlogCommendClickRateRank_13,index_13"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/u011747846/article/details/80602786" target="_blank" title="Center-Loss - 當凌絕頂,俯瞰天下">
<h4 class="text-truncate oneline" style="width: 610px;">
Center-Loss - 當凌絕頂,俯瞰天下 </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/E/E/D/3_u011747846.jpg" alt="u011747846" class="avatar-pic">
<span class="namebox">
<span class="name">u011747846</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">06-07</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
1020</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/u011747846/article/details/80602786" target="_blank" title="Center-Loss - 當凌絕頂,俯瞰天下">
<span class="desc oneline">摘要
這篇部落格是對論文A Discriminative Feature Learning Approach for Deep Face Recognition的總結。這篇論文中，作者提出了一種新的…
來自： 當凌絕頂,俯瞰天下
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/u011808673/article/details/80491361,BlogCommendClickRateRank_14,index_14"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/u011808673/article/details/80491361,BlogCommendClickRateRank_14,index_14"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/u011808673/article/details/80491361" target="_blank" title="人臉識別-損失函式改進方法之L-softmax loss center-loss A-softmax loss focal loss - Nine days">
<h4 class="text-truncate oneline" style="width: 618px;">
<em>人臉</em>識別-損失函式改進方法之L-softmax loss center-loss A-softmax loss focal loss - Nine days </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/6/7/3/3_u011808673.jpg" alt="u011808673" class="avatar-pic">
<span class="namebox">
<span class="name">u011808673</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">05-29</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
708</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/u011808673/article/details/80491361" target="_blank" title="人臉識別-損失函式改進方法之L-softmax loss center-loss A-softmax loss focal loss - Nine days">
<span class="desc oneline">
而ArcFac中的loss與其他loss不同之處在於下圖：
與CosFace的解釋相同，只是ArcFace…
來自： Nine days
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/intflojx/article/details/82455560,BlogCommendClickRateRank_15,index_15"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/intflojx/article/details/82455560,BlogCommendClickRateRank_15,index_15"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/intflojx/article/details/82455560" target="_blank" title="COCO loss （人臉識別損失函式） - intflojx的部落格">
<h4 class="text-truncate oneline" style="width: 618px;">
COCO loss （<em>人臉</em>識別損失函式） - intflojx的部落格 </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/6/5/B/3_intflojx.jpg" alt="intflojx" class="avatar-pic">
<span class="namebox">
<span class="name">intflojx</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">09-06</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
129</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/intflojx/article/details/82455560" target="_blank" title="COCO loss （人臉識別損失函式） - intflojx的部落格">
<span class="desc oneline"> 2017年nips的一篇做分類和識別的工作，其中在人臉識別任務上也做了實驗，Rethinking Feature Discrimination and Polymerization for Larg...</span>
</a>
<span class="blog_title_box oneline"><a target="_blank" href="https://blog.csdn.net/intflojx">來自： <span class="blog_title"> intflojx的部落格</span></a></span>
</p>
</div>
</div>
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/qq_41185868/article/details/82929877,BlogCommendClickRateRank_16,index_16"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/qq_41185868/article/details/82929877,BlogCommendClickRateRank_16,index_16"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/qq_41185868/article/details/82929877" target="_blank" title="CV之FRec之Loss：人臉識別中常用的幾種損失簡介、使用方法之詳細攻略 - 一個處女座的程式猿">
<h4 class="text-truncate oneline" style="width: 626px;">
CV之FRec之Loss：<em>人臉</em>識別中常用的幾種損失簡介、使用方法之詳細攻略 - 一個處女座的程式猿 </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/B/4/2/3_qq_41185868.jpg" alt="qq_41185868" class="avatar-pic">
<span class="namebox">
<span class="name">qq_41185868</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">10-03</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
77</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/qq_41185868/article/details/82929877" target="_blank" title="CV之FRec之Loss：人臉識別中常用的幾種損失簡介、使用方法之詳細攻略 - 一個處女座的程式猿">
<span class="desc oneline">CV之FRec之Loss：人臉識別中常用的幾種損失簡介、使用方法之詳細攻略
T1、Triplet Loss
《FaceNet: A Unified Embedding for Face Re…
來自： 一個處女座的程式猿
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/Iriving_shu/article/details/77822201,BlogCommendClickRateRank_17,index_17"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/Iriving_shu/article/details/77822201,BlogCommendClickRateRank_17,index_17"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/Iriving_shu/article/details/77822201" target="_blank" title="人臉識別系列-Centerloss論文閱讀筆記 - Iriving_shu的部落格">
<h4 class="text-truncate oneline" style="width: 618px;">
<em>人臉</em>識別系列-<em>Centerloss</em>論文閱讀筆記 - Iriving_shu的部落格 </h4>
<div class="info-box d-flex align-content-center">
<!-- <p class="avatar">
<img src="https://avatar.csdn.net/B/3/9/3_iriving_shu.jpg" alt="Iriving_shu" class="avatar-pic">
<span class="namebox">
<span class="name">Iriving_shu</span>
<span class="triangle"></span>
</span>
</p> -->
<p class="date-and-readNum">
<span class="date hover-show">09-03</span>
<span class="read-num hover-hide">
<svg class="icon csdnc-yuedushu" aria-hidden="true">
<use xlink:href="#csdnc-m-passwords-visible"></use>
</svg>
876</span>
</p>
</div>
</a>
<p class="content" style="width: 702px;">
<a href="https://blog.csdn.net/Iriving_shu/article/details/77822201" target="_blank" title="人臉識別系列-Centerloss論文閱讀筆記 - Iriving_shu的部落格">
<span class="desc oneline">好記心不如爛筆頭，以此記錄自己的學習歷程。
這篇論文 Centerloss來自Eccv2016。本文主要講解其核心原理以及原始碼和實驗。
論文地址：https://ydwen.github.io/p…
來自： Iriving_shu的部落格
<div class="recommend-item-box recommend-box-ident type_blog clearfix" data-track-view="{"mod":"popu_614","con":",https://blog.csdn.net/u014380165/article/details/76946339,BlogCommendClickRateRank_18,index_18"}" data-track-click="{"mod":"popu_614","con":",https://blog.csdn.net/u014380165/article/details/76946339,BlogCommendClickRateRank_18,index_18"}" data-flg="true">
<div class="content" style="width: 702px;">
<a href="https://blog.csdn.net/u0143