【论文笔记】Object Contour Detection with a Fully Convolutional Encoder-Decoder Network

Object Boundary Extraction Based on a Fully Convolutional Encoder-Decoder Network

---

使用卷积编码解码网络检测主要目标的边缘

网络结构是：
编码：VGG-16
解码：反池化-卷积-激活-dropout

卷积核:

The number of channels in each decoder layer is meticulously planned to enable unpooling operations within their respective corresponding max pooling layers.

dropout:

Incorporating a dropout layer post each ReLU layer within the decoder, as described in Section 3. The complete configuration of the decoder network is detailed in Table 1.

损失函数：

loss function is simply the pixel-wise logistic loss.

反池化：

We parameterize our encoder based on the VGG-16 network [48], which is configured up to the “fc6” layer. Our decoder is built through an alternating process that combines unpooling and convolutional layers, wherein each unpooling layer reutilizes switches from corresponding max-pooling layers in the encoder to upscale feature maps towards achieving dense predictions of image size.

反池化细节：
<>

---

During the training process, we choose to freeze the encoder weights, specifically using a VGG-16 architecture, and exclusively focus on optimizing the decoder network.

---