![深度学习之模型设计:核心算法与案例实践](https://wfqqreader-1252317822.image.myqcloud.com/cover/822/33114822/b_33114822.jpg)
上QQ阅读APP看本书,新人免费读10天
设备和账号都新为新人
![](https://epubservercos.yuewen.com/60C819/17725770607802806/epubprivate/OEBPS/Images/39030_2_1.jpg?sign=1739298111-wnK4FRCyC17G8juvdsicTXZbavN8laUe-0-3a6ed9a68db28badf5ff17c4b7361d1d)
图1.7 灰度图与彩色图
![](https://epubservercos.yuewen.com/60C819/17725770607802806/epubprivate/OEBPS/Images/39030_2_2.jpg?sign=1739298111-nyyAkcUpSdiVzWE6Vsv1o0ZS7TtfmwE4-0-640f30940adb5684e2d097f3ffa429cf)
图1.8 灰度图的直方图与彩色图的直方图
![](https://epubservercos.yuewen.com/60C819/17725770607802806/epubprivate/OEBPS/Images/39030_2_3.jpg?sign=1739298111-FuSSQ6oAE4duBGkRB1brslGkhCXhlKIx-0-3b7f9dab1778b4dac396e5cf99a6e6b8)
图2.15 基于动量项的SGD示意
![](https://epubservercos.yuewen.com/60C819/17725770607802806/epubprivate/OEBPS/Images/39030_2_4.jpg?sign=1739298111-cEyJ53lZ8WZy41CdaGcyHlCismTdCLkM-0-1723b1e5861dc43d45f0569e1f7de80c)
图4.3 TDNN示意
![](https://epubservercos.yuewen.com/60C819/17725770607802806/epubprivate/OEBPS/Images/39030_3_1.jpg?sign=1739298111-Sq2vTnuKVl58BdxLkLuPI692C7TbRCB2-0-f927c4fb515b1761e1fed082ae99546d)
图6.1 AlexNet第一个卷积层的96个通道的可视化结果
![](https://epubservercos.yuewen.com/60C819/17725770607802806/epubprivate/OEBPS/Images/39030_3_2.jpg?sign=1739298111-86dXEf0YtVTPsM2M0Ydsi2ncIeAwowxK-0-b2fd079cee004a6a06a8a3448095c902)
图7.13 Allconv5_SEG训练结果
![](https://epubservercos.yuewen.com/60C819/17725770607802806/epubprivate/OEBPS/Images/39030_3_3.jpg?sign=1739298111-hcvHgiR7UoUhBLFvPs67hbBLYcZhgogT-0-505048e00771bd738e3601bd55cfcc62)
图7.14 Allconv5_SEG使用224×224的分辨率进行测试的分割结果
![](https://epubservercos.yuewen.com/60C819/17725770607802806/epubprivate/OEBPS/Images/39030_3_4.jpg?sign=1739298111-e7qJqik5M1gQJnF8ISDM4qVCw0drPj6B-0-9ccc44683b0800ae511082ff9a8b4d75)
图7.16 Allconv5_SEG与Allconv5_Skip_SEG的训练结果
![](https://epubservercos.yuewen.com/60C819/17725770607802806/epubprivate/OEBPS/Images/39030_4_1.jpg?sign=1739298111-ktoVzixoA29xW2XxqNvHDOrsqwuKHTwG-0-f3a3fade918da2ea9b52a71805dabb8c)
图7.17 Allconv5_Skip_SEG使用224×224的分辨率进行测试的分割结果
![](https://epubservercos.yuewen.com/60C819/17725770607802806/epubprivate/OEBPS/Images/39030_4_2.jpg?sign=1739298111-BL74Kh9IjZ1NzZGCz52YadrUlOZuaCFf-0-139270f60b219b6bcc09a33fe44b6811)
图7.18 Allconv5_SEG与Allconv5_Skip_SEG使用448×448的分辨率进行测试的分割结果
![](https://epubservercos.yuewen.com/60C819/17725770607802806/epubprivate/OEBPS/Images/39030_4_3.jpg?sign=1739298111-uqsSC6ofremDo7gqKCIlRxD3DtzZIenS-0-9c60679e36ee5591c8b252ddcec8bd40)
图8.11 嘴唇图像与标注示意
![](https://epubservercos.yuewen.com/60C819/17725770607802806/epubprivate/OEBPS/Images/39030_4_4.jpg?sign=1739298111-QloUhel7oEZg2n7ypCNwsC4hJJttDkNu-0-10e99dffb02ac7c9e5b088a954201dd9)
图8.13 MobileSegNet_MOUTH160精度曲线和损失曲线
![](https://epubservercos.yuewen.com/60C819/17725770607802806/epubprivate/OEBPS/Images/39030_5_1.jpg?sign=1739298111-LUmhyxvmWIAGngcWS4PfQhSevcf7KMCv-0-d5d75390e84e85c206c6f823ee6e0296)
图9.16 可变形卷积的感受野示意(使用大小为3×3的卷积核)
![](https://epubservercos.yuewen.com/60C819/17725770607802806/epubprivate/OEBPS/Images/39030_5_2.jpg?sign=1739298111-jJbvsTe9RIIyaOluNoI3bF6ftR9IWWTv-0-b5604ae87deda1e550ec36ecc2ac8564)
图12.3 简单的三维卷积网络
![](https://epubservercos.yuewen.com/60C819/17725770607802806/epubprivate/OEBPS/Images/39030_5_3.jpg?sign=1739298111-UXNVabpgNWFR4uUtDmZBx7wQpfaPa6hQ-0-2be3116675a82a4e602c58eb071fe5b7)
图12.12 不同比例下的训练集和测试集精度