Why CNN for Image? Feature Extraction with Convolution Kernel This“patchy”operation is convolution Filter of size 4x4:16 different weights 1)Apply a set of weights-a filter-to extract local features Apply this same filter to 4x4 patches in input 2)Use multiple filters to extract different features -Shift by 2 pixels for next patch 3)Spatially share parameters of each filter Hangchou①ianti Universi的杭州电子科技大学 School of Computer Science and Technology计算机学院周文库
Hangzhou Dianzi University 杭州电子科技大学 School of Computer Science and Technology 计算机学院 周文晖 - Filter of size 4×4 : 16 different weights - Apply this same filter to 4×4 patches in input - Shift by 2 pixels for next patch This “patchy” operation is convolution 1) Apply a set of weights – a filter – to extract local features 2) Use multiple filters to extract different features 3) Spatially share parameters of each filter Why CNN for Image？ Feature Extraction with Convolution Kernel
Why CNN for Image? Convolution Operator 以I为中心 的邻域区域 Kernel 41421 hh2 h3 14因6 米 ns hs hs 1,18L I7 Iis 1g 卷积计算 I5=hI1+hI2+hI3+h。I4+hI5+h4I6+h3·I,+h2·Ig+h·Ig 由于模板通常都是中心对称的，即可忽略模板以中心反转的过程，有 Ig=h·L1+h2I2+hI3+h4I4+hL+h。I6+hI+h。Ig+hIg Hangzhou①ianzi Universi的杭州电子科技大学 School of Computer Science and Tecfnology计算机学院周文库
Hangzhou Dianzi University 杭州电子科技大学 School of Computer Science and Technology 计算机学院 周文晖 Why CNN for Image？ Convolution Operator I1 I2 I3 I4 I5 I6 I7 I8 I9 h1 h2 h3 h4 h5 h6 * h7 h8 h9 5 91 82 73 64 55 46 37 28 19 I hI hI hI hI hI hI hI hI hI 由于模板通常都是中心对称的，即可忽略模板以中心反转的过程，有 5 11 2 2 3 3 4 4 5 5 6 6 7 7 8 8 9 9 I hI hI hI hI hI hI hI hI hI 以 I5 为中心 的邻域区域 卷积计算 Kernel
Input Yolue (tpad 1)(7z7x3) Filter W (3x3x3) Filter W1 (3x3x3) Output Volune (3x3x2) Why CNN for Image? x,:,01 w0[:,:,01 w1I,,0] g,,0] 0000000 11-1 -1-10 ▣°3 0011220 -101 -110 611 0011000 10 -110 4-31 0110100 w0[:,:,1 m1I:,:,1] a「：，：，1】 Convolution Operator 0101110 可四网 1-10 -164 0020100 00可 -10-1 -234 000000 1▣0 -100 -1-3-3 w0[:2 w1I:,,21 1000000 610 -101 0 11200 101 082 1128 0▣1 0-10 0120020 Bias bo (1:11) Bias bl (1x1x1) 0021/210 b1I:,:,0] 020129/0 01 090060 克▣.2 toggle novenent 000600/0 0202920 00y210 0102210 02 02000 0001120 0000000 卷积滤波过程：遍历图像中所有像素，计算每个像素的邻域与模板的卷积值。 http://blog.cadn.net/Jesse Mx Hangchou Dianzi Universi的y杭州电子科技大学 School of Computer Science and Technology计算机学院周文库
Hangzhou Dianzi University 杭州电子科技大学 School of Computer Science and Technology 计算机学院 周文晖 Why CNN for Image？ Convolution Operator F F ′ 卷积滤波过程：遍历图像中所有像素，计算每个像素的邻域与模板的卷积值
Why CNN for Image? Convolution Operator Suppose we want to compute the convolution of a 5x5 image and a 3x3 filter: 1 1 10 0 0 1 1 1 0 101 0 0 1 0 0 1 0 01 0 110 filter image filter feature map We slide the 3x3 filter over the input image,element-wise multiply,and add the outputs... Hangzhou①ianzi Universi的抗州电子科技大学 School of Computer Science andT2 chnology计算机学院周文库
Hangzhou Dianzi University 杭州电子科技大学 School of Computer Science and Technology 计算机学院 周文晖 image filter Suppose we want to compute the convolution of a 5x5 image and a 3x3 filter: We slide the 3x3 filter over the input image, element-wise multiply, and add the outputs… Why CNN for Image？ Convolution Operator
Why CNN for Image? Feature Extraction with Convolution 0-10 010 121 151 41 0o6 0 010 121 Original Sharpen Edge Detect “Strong”Edge Detect Hangzhou①ianzi Universi的杭州电子科技大学 School of Computer Science and Tecfnology计算机学院周文库
Hangzhou Dianzi University 杭州电子科技大学 School of Computer Science and Technology 计算机学院 周文晖 Original Sharpen Edge Detect “Strong” Edge Detect Why CNN for Image？ Feature Extraction with Convolution