Table 5 Performance impact of filters and histogram equalization on the proposed framework across multiple deep learning architectures for test set (Results in bold highlight the best accuracy value in percentage for each network architecture).

From: A multi-filter deep transfer learning framework for image-based autism spectrum disorder detection

Filters (\(\mathcal {F}\)) / Structure

AffectNet

AlexNet

ResNet-50

ViTSwin

VGG16

VGG19

ViT

ViTFER

None

87.67

79.33

75.67

90.33

73.00

72.00

87.67

87.00

high

89.33

77.33

69.67

89.33

68.33

62.00

82.00

80.67

smooth

90.33

76.33

71.67

88.67

71.67

71.33

88.33

88.33

original

91.00

78.67

74.00

91.33

71.67

72.00

89.00

87.33

histeq_high

89.33

77.33

69.00

89.00

69.00

63.33

81.67

81.00

histeq_smooth

89.33

78.00

73.33

89.33

71.67

72.67

87.33

86.00

histeq_original

88.67

79.67

73.67

90.67

73.00

71.00

89.00

88.33

high+histeq_high

88.67

74.00

67.00

89.00

67.67

60.33

81.00

80.00

original+histeq_high

90.00

79.33

75.33

90.67

73.00

71.00

89.00

86.00

original+smooth+high

90.67

78.33

74.33

90.33

74.00

72.00

89.33

87.00

smooth+histeq_smooth

89.33

77.33

72.33

90.00

71.33

74.00

87.33

88.33

original+histeq_smooth

89.33

78.67

72.67

92.67

70.00

71.33

87.33

87.67

original+histeq_original

88.67

79.00

77.00

90.33

72.00

71.00

89.00

86.33

histeq_original+smooth+high

89.67

78.33

76.67

91.67

72.67

72.33

90.67

86.33

histeq_original+histeq_high

89.00

79.67

75.67

91.33

71.00

69.33

89.00

85.67

histeq_original+histeq_smooth

88.33

79.00

74.00

91.67

73.67

70.67

90.00

87.00

original+histeq_smooth+histeq_high

90.33

79.33

74.33

91.33

73.00

72.33

89.00

86.67

histeq_original+histeq_smooth+histeq_high

89.67

78.67

74.67

92.00

72.00

72.67

89.00

85.00

original+smooth+high+histeq_original+histeq_smooth+histeq_high

90.00

81.67

75.67

91.00

73.67

72.33

89.67

87.33

  1. Significant values are in bold.