SENet(仅供参考学习转自大神博主）

最新推荐文章于 2024-07-26 00:40:46 发布

樨潮

最新推荐文章于 2024-07-26 00:40:46 发布

阅读量525

点赞数

分类专栏：分类网络

分类网络专栏收录该内容

13 篇文章

订阅专栏

SENet（Squeeze-and-Excitation Networks）是Momenta在ILSVRC 2017分类任务中夺冠的模型，其Top-5错误率降至2.251%，相比2016年冠军提升了约25%的准确性。该文详细介绍了SENet的整体结构，包括Ftr、Fsq、Fex、Fscale等关键函数，以及SE模块的设计原理。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

                        
				版权声明：本文为博主原创文章，未经博主允许不得转载。					https://blog.youkuaiyun.com/qq_14845119/article/details/80888940				</div>
							<link rel="stylesheet" href="https://csdnimg.cn/release/phoenix/template/css/ck_htmledit_views-f57960eb32.css">
							            <link rel="stylesheet" href="https://csdnimg.cn/release/phoenix/template/css/ck_htmledit_views-f57960eb32.css">
					<div class="htmledit_views" id="content_views">
            <p>论文：<a href="https://arxiv.org/abs/1709.01507" rel="nofollow" target="_blank">Squeeze-and-Excitation Networks</a></p><p>Github<span style="font-family:'宋体';">：</span><span style="font-family:Calibri;"><a href="https://github.com/hujie-frank/SENet" rel="nofollow" target="_blank">https://github.com/hujie-frank/SENet</a></span></p><p>&nbsp;</p><p>Momenta ILSVRC 2017 <span style="font-family:'宋体';">分类任务的冠军，</span>top-5 错误达到2:251% ，<span style="font-family:'宋体';">相比</span>2016<span style="font-family:'宋体';">年冠军有</span>∼<span style="font-family:Calibri;">25% </span>的准确性提升。</p><p>&nbsp;</p><p><span style="color:rgb(255,0,0);">模型整体结构：</span></p><p style="text-align:center;"><span style="color:rgb(255,0,0);"><img src="https://img-blog.youkuaiyun.com/20180702194450490?watermark/2/text/aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzE0ODQ1MTE5/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70" alt=""><br></span></p><p></p><p>Ftr<span style="font-family:'宋体';">函数表示一系列的卷积，池化操作等。其中</span><span style="font-family:Calibri;">X</span><span style="font-family:'宋体';">代表输入图片，</span><span style="font-family:Calibri;">v</span><span style="font-family:'宋体';">代表卷积核。</span></p><p style="text-align:center;"><img src="https://img-blog.youkuaiyun.com/2018070219452627?watermark/2/text/aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzE0ODQ1MTE5/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70" alt=""></p><p style="text-align:center;"><img src="https://img-blog.youkuaiyun.com/20180702194749503?watermark/2/text/aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzE0ODQ1MTE5/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70" alt=""><br></p><p></p><p>Fsq<span style="font-family:'宋体';">函数代表</span><span style="font-family:Calibri;">global average pooling</span><span style="font-family:'宋体';">，</span></p><p style="text-align:center;"><img src="https://img-blog.youkuaiyun.com/20180702194815993?watermark/2/text/aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzE0ODQ1MTE5/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70" alt=""></p><p></p><p>Fex<span style="font-family:'宋体';">函数代表卷积和激活操作，具体的说，首先是基于</span><span style="font-family:Calibri;">r=16</span><span style="font-family:'宋体';">系数的降维操作，并且在降维后做了</span><span style="font-family:Calibri;">Relu</span><span style="font-family:'宋体';">操作，然后基于</span><span style="font-family:Calibri;">r=16</span><span style="font-family:'宋体';">系数的升维操作，生维后做</span><span style="font-family:Calibri;">sigmoid</span><span style="font-family:'宋体';">激活操作。</span></p><p style="text-align:center;"><img src="https://img-blog.youkuaiyun.com/20180702194543160?watermark/2/text/aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzE0ODQ1MTE5/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70" alt=""><br></p><p style="text-align:center;"><img src="https://img-blog.youkuaiyun.com/20180702194833201?watermark/2/text/aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzE0ODQ1MTE5/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70" alt=""><br></p><p></p><p>Fscale<span style="font-family:'宋体';">代表基于</span><span style="font-family:Calibri;">channel</span><span style="font-family:'宋体';">的</span><span style="font-family:Calibri;">elementwise</span><span style="font-family:'宋体';">乘积运算。</span></p><p style="text-align:center;"><img src="https://img-blog.youkuaiyun.com/2018070219485476?watermark/2/text/aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzE0ODQ1MTE5/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70" alt=""><br></p><p></p><p><span style="color:rgb(255,0,0);"><br></span></p><p><span style="color:rgb(255,0,0);">SE<span style="font-family:'宋体';">模块：</span></span></p><p style="text-align:center;"><span style="color:rgb(255,0,0);"><span style="font-family:'宋体';"><img src="https://img-blog.youkuaiyun.com/20180702194918177?watermark/2/text/aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzE0ODQ1MTE5/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70" alt=""><br></span></span></p><p><span style="color:rgb(255,0,0);"><span style="font-family:'宋体';"></span></span></p><p><span style="font-family:'宋体';">上图表示了基于</span>Inception<span style="font-family:'宋体';">系列网络的</span><span style="font-family:Calibri;">SE</span><span style="font-family:'宋体';">模块设计和基于</span><span style="font-family:Calibri;">Resnet</span><span style="font-family:'宋体';">这种带</span><span style="font-family:Calibri;">shortcut</span><span style="font-family:'宋体';">结构的</span><span style="font-family:Calibri;">SE</span><span style="font-family:'宋体';">模块设计。其中压缩比</span><span style="font-family:Calibri;">r=16</span><span style="font-family:'宋体';">时候，获得了参数量和速度的最佳平衡。</span></p><br><br><p><br></p>            </div>
                  </div>