Image Compression Based on Hybrid Domain Attention and Postprocessing Enhancement

Table 1

The proposed method's detailed parameter settings. “CONV” signifies convolutional layer. “2↓” signifies downsampling in the stride of 2. “2↑” signifies upsampling in the stride of 2. HDAM represents the hybrid domain attention module.

Encoder	Decoder	Hyperprior encoder	Hyperprior decoder	Context model
CONV:5 × 5 × 128 2↓	DECONV:5 × 5 × 128 2↑	CONV:3 × 3 × 128	DECONV:5 × 5 × 192 2↑	Masked:5 × 5 × 5 × 24
GDN	IGDN	Leaky ReLU	Leaky ReLU	CONV:1 × 1 × 48
CONV:5 × 5 × 128 2↓	DECONV:5 × 5 × 128 2↑	HDAM	HDAM	ReLU
GDN	IGDN	CONV:5 × 5 × 128 2↓	DECONV:5 × 5 × 288 2↑	CONV:1 × 1 × 96
HDAM	HDAM	Leaky ReLU	Leaky ReLU	ReLU
CONV:5 × 5 × 128 2↓	DECONV:5 × 5 × 128 2↑	CONV:5 × 5 × 192 2↓	DECONV:3 × 3 × 384 2↑	CONV:1 × 1 × 1 × 2
GDN	GDN	HDAM
CONV:5 × 5 × 192 2↓	DECONV:5 × 5 × 3 2↑
HDAM	HDAM		HDAM