• Laser & Optoelectronics Progress
  • Vol. 56, Issue 23, 231002 (2019)
Xiaojia Jiang and Shuhui Gao*
Author Affiliations
  • Institute of Forensic Science, People's Public Security University of China, Beijing 102623, China
  • show less
    DOI: 10.3788/LOP56.231002 Cite this Article Set citation alerts
    Xiaojia Jiang, Shuhui Gao. Automatic Classification of Microscopic Hair Images Based on Improved Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2019, 56(23): 231002 Copy Citation Text show less
    Basic structure of convolutional neural network
    Fig. 1. Basic structure of convolutional neural network
    Examples of sample dataset images
    Fig. 2. Examples of sample dataset images
    Structural design of Hair-Net
    Fig. 3. Structural design of Hair-Net
    Flow chart of automatic classification for microscopic hair images
    Fig. 4. Flow chart of automatic classification for microscopic hair images
    Comparison of Hair-Net classification accuracy with and without crop_size
    Fig. 5. Comparison of Hair-Net classification accuracy with and without crop_size
    Comparison of Hair-Net loss with and without crop_size
    Fig. 6. Comparison of Hair-Net loss with and without crop_size
    Classification accuracy under optimal network performance
    Fig. 7. Classification accuracy under optimal network performance
    Loss under optimal network performance
    Fig. 8. Loss under optimal network performance
    Software/hardwareExperimental configuration
    CPUIntel(R)4 Core(TM)i5-7200U,2.50 GHz
    GPUNVIDIA GeForce 940MX
    RAM8.00 GB
    FrameCaffe
    ToolsoftwareMatlab2016a, Python2.7, VisualStudio 2013, Adobe Photoshop CS4
    Table 1. Experimental environment configuration and parameters
    LayerKernel sizeStridesPadOutput sizeOutput number
    Input223×2233
    Convolution1+ ReLU19*94155×5596
    Max pooling13*32027×2796
    Convolution2+ReLU23*31127×27256
    Convolution3+ReLU33*31127×27256
    Max pooling33*32013×13256
    Convolution4+ReLU43*31113×13384
    Convolution5+ ReLU53*31113×13384
    Convolution6+ReLU63*31113×13256
    Max pooling63*3206×6256
    Fully connected7+ ReLU7+ Dropout71×14096
    Fully connected8+ ReLU8+ Dropout81×11024
    Fully connected91×16
    Output6
    Table 2. Configuration parameters of Hair-Net
    Batch_sizeTrain accuracy /%LossTest accuracy /%
    1676.151.14972.43
    3283.420.82280.97
    6488.500.70986.39
    12894.810.38893.66
    256ErrorErrorError
    Table 3. Effect of batch size on experimental results
    ModelstructureTrainaccuracy /%LossTestaccuracy /%
    Hair-Net96.150.24795.32
    Hair-Net+LRN95.490.31694.57
    Hair-Net+BN97.820.19997.16
    Table 4. Effects of LRN and BN on experimental results
    Xiaojia Jiang, Shuhui Gao. Automatic Classification of Microscopic Hair Images Based on Improved Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2019, 56(23): 231002
    Download Citation