Table 6.

Testing results on four datasets for network structures with different pooling layers. Only the results for the L₃ network are presented here for clearer comparisons of different network structures.

Algorithms		Avg Pooling	Max Pooling	Gaussian Pooling	Gaussian Smoothing
		L = 3	L = 3	L = 3	L = 3
Kodak24	R	40.95	40.74	40.86	41.22
	G	44.09	43.86	44.01	44.34
	B	40.35	40.11	40.28	40.59
	RGB	41.47	41.25	41.39	41.72
McMaster	R	37.83	37.79	37.71	38.01
	G	40.58	40.53	40.56	40.74
	B	36.07	36.03	35.97	36.33
	RGB	37.70	37.66	37.61	37.91
Urban100	R	36.67	36.34	36.58	36.84
	G	40.29	39.99	40.22	40.46
	B	36.68	36.32	36.56	36.90
	RGB	37.51	37.17	37.41	37.70
Manga109	R	36.93	36.72	36.90	37.27
	G	41.53	41.30	41.54	41.93
	B	36.75	36.54	36.74	37.01
	RGB	37.86	37.65	37.84	38.17
Ave.	R	38.10	37.90	38.01	38.33
	G	41.62	41.42	41.58	41.87
	B	37.46	37.25	37.39	37.71
	RGB	38.64	38.43	38.56	38.88

Where Avg pooling, Max pooling, Gaussian pooling, and Gaussian smoothing denote 2 × 2 average pooling, 2 × 2 max pooling, Gaussian smoothing followed by 2 × 2 down-sampling, and Gaussian smoothing layer (used in this paper), respectively. Since pooling changes the image size, in each image reconstruction node in the original network, we replace the 1 × 1 convolution with up-sampling implemented by a transposed convolution. It can be seen that the network adopting Gaussian smoothing layers can extract image features more efficiently than other pooling approaches, thus obtaining a little better accuracy due to its capability to extract image features through multi-scale receptive fields and its preservation of entire image information.