Table 1.

Proposed CNN architecture layer information. CNN-1,2,3 have 3-channel inputs. CNN-4 has 1,3, and 4-channel inputs. To simplify the table, dropout, pooling, ReLU layers are not listed

Architecture	Layer	Kernel	Filter	Output
CNN-1,2,3 (Facial Component Segmentation Input: 16 × 16 blocks)	conv1	5 × 5	16	16×16×16
	conv2	5 × 5	16	8×8×16
	conv3	5 × 5	32	4×4×32
	conv4	4 × 4	32	1×1×32
CNN-4 (Facial Expr. Recognition Input: 64 × 64 full image)	conv1	5 × 5	64	64×64×64
	conv2	5 × 5	32	32×32×32
	conv3	5 × 5	32	16×16×32
	conv4	5 × 5	64	8×8×64
	conv5	4 × 4	64	1×1×64