Chapter 5 Convolutional Neural Network (CNN)

1. 5.1 Constitutive layers of convolutional neural networks 170

2. 5.2 How does convolution detect edge information? 171

3. 5.2 Several basic definitions of convolution? 174

4. 5.2.1 Convolution kernel size 174

5. 5.2.2 Step size of the convolution kernel 174

6. 5.2.3 Edge Filling 174

7. 5.2.4 Input and Output Channels 174

8. 5.3 Convolution network type classification? 174

9. 5.3.1 Ordinary Convolution 174

10. 5.3.2 Expansion Convolution 175

11. 5.3.3 Transposition Convolution 176

12. 5.3.4 Separable Convolution 177

13. 5.3 Schematic of 12 different types of 2D convolution? 178

14. 5.4 What is the difference between 2D convolution and 3D convolution? 181

15. 5.4.1 2D Convolution 181

16. 5.4.2 3D Convolution 182

17. 5.5 What are pooling methods? 183

18. 5.5.1 General Pooling 183

19. 5.5.2 Overlapping Pooling (OverlappingPooling) 184

20. 5.5.3 Spatial Pyramid Pooling 184

21. 5.6 1x1 convolution? 186

22. 5.7 What is the difference between the convolutional layer and the pooled layer? 187

23. 5.8 Does a larger convolution kernel, improve the kernel? 189

24. 5.9 Can each convolution use only one size of convolution kernel? 189

25. 5.10 How can I reduce the amount of convolutional parameters? 190

26. 5.11 Must convolution operations consider both channels and zones? 191

27. 5.12 What are the benefits of using wide convolution? 192

28. 5.12.1 Narrow Convolution and Wide Convolution 192

29. 5.12.2 Why use wide convolution? 192

30. 5.13 Which depth of the convolutional layer output is the same as the number of parts? 192

31. 5.14 How do I get the depth of the convolutional layer output? 193

32. 5.15 Is the activation function usually placed after the operation of the convolutional neural network? 194

33. 5.16 How do you understand that the maximum pooling layer is a little smaller? 194

34. 5.17 Understanding Image Convolution and Deconvolution 194

35. 5.17.1 Image Convolution 194

36. 5.17.2 Image Deconvolution 196

37. 5.18 Image Size Calculation after Different Convolutions? 198

38. 5.18.1 Type division 198

39. 5.18.2 Calculation formula 199

40. 5.19 Step size, fill size and input and output relationship summary? 199

41. 5.19.1 No 0 padding, unit step size 200

42. 5.19.2 Zero fill, unit step size 200

43. 5.19.3 Not filled, non-unit step size 202

44. 5.19.4 Zero padding, non-unit step size 202

45. 5.20 Understanding deconvolution and checkerboard effects 204

46. 5.20.1 Why does the board phenomenon appear? 204

47. 5.20.2 What methods can avoid the checkerboard effect? 205

48. 5.21 CNN main calculation bottleneck? 207

49. 5.22 CNN parameter experience setting 207

50. 5.23 Summary of methods for improving generalization ability 208

51. 5.23.1 Main methods 208

52. 5.23.2 Experimental proof 208

53. 5.24 What are the connections and differences between CNN and CLP? 213

54. 5.24.1 Contact 213

55. 5.24.2 Differences 213

56. 5.25 Does CNN highlight commonality? 213

57. 5.25.1 Local connection 213

58. 5.25.2 Weight sharing 214

59. 5.25.3 Pooling Operations 215

60. 5.26 Similarities and differences between full convolution and Local-Conv 215

61. 5.27 Example Understanding the Role of Local-Conv 215

62. 5.28 Brief History of Convolutional Neural Networks 216

P