Table 2 The architecture of CNN-based image extractor.

From: Fully connected-convolutional (FC-CNN) neural network based on hyperspectral images for rapid identification of P. ginseng growth years

Layer

Layer shape (output shape)

Weights number

Input

(None, 256, 1024, 6*)

0

Conv

(None, 256, 1024, 64)

3520

BatchNorm

(None, 256, 1024, 64)

256

MaxPooling

(None, 128, 512, 64)

0

Conv

(None, 128, 512, 128)

73,856

BatchNorm

(None, 128, 512, 128)

512

MaxPooling

(None, 64, 256, 128)

0

Conv

(None, 64, 256, 128)

147,584

BatchNorm

(None, 64, 256, 128)

512

MaxPooling

(None, 32, 128, 128)

0

Conv

(None, 32, 128, 64)

73,792

BatchNorm

(None, 32, 128, 64)

256

MaxPooling

(None, 16, 64, 64)

0

Conv

(None, 16, 64, 32)

18,464

BatchNorm

(None, 16, 64, 32)

128

MaxPooling

(None, 8, 32, 32)

0

Flatten

(None, 8192)

0

Dropout

(None, 8192)

0

Dense

(None, 512)

4,194,816

  1. *The number of input images was tried in 0–6. Here only show the case that the CNN used maximum 6 images of different spectral bands.
Back to article page