segmentation_models - mobilenet-unet

[junxnone]

Arch

Backbone-mobilenet + Decoder x 5 + Final Conv + Activation

• 先提取 Shape/(2^5) 的 feature map
• 再 upsample 2^5 到原 resolution

Input	448x448x3 - HxWxC
Output	448x448xclasses_num or 448x448x(classes_num+1) - H/(2^5) x W/(2^5) x classes_num

Backbone

backbone	mobilenet
Input	448x448x3
Output	14x14x1024

Decoder

Stage	filter	Transpose Input	Transpose Output	Concat Layer	Concat Input	Concat Output	Output
0	64	14x14x1024	28x28x64	conv_pw_11_relu	28x28x512	28x28x576	28x28x64
1	32	28x28x64	56x56x32	conv_pw_5_relu	56x56x256	56x56x288	56x56x32
2	16	56x56x32	112x112x16	conv_pw_3_relu	112x112x128	112x112x144	112x112x16
3	8	112x112x16	224x224x8	conv_pw_1_relu	224x224x64	224x224x72	224x224x8
4	4	224x224x8	448x448x4	None	None	None	448x448x4

Final Conv + Activation

Input	448x448x4
Output	448x448xclasses_num or 448x448x(classes_num+1)

• Activation
  • sigmoid
  • softmax

[junxnone]

672