• Acta Optica Sinica
  • Vol. 40, Issue 8, 0811005 (2020)
Lu Jin1、2、3, Shijian Liu1、3, Xiao Wang1、2、3, and Fanming Li1、3、*
Author Affiliations
  • 1Shanghai Institute of Technical Physics, Chinese Academy of Sciences, Shanghai 200083, China
  • 2University of Chinese Academy of Sciences, Beijing 100049, China
  • 3Key Laboratory of Infrared System Detection and Imaging Technology, Chinese Academy of Sciences, Shanghai 200083, China
  • show less
    DOI: 10.3788/AOS202040.0811005 Cite this Article Set citation alerts
    Lu Jin, Shijian Liu, Xiao Wang, Fanming Li. Infrared Aircraft Classification Method with Small Samples Based on Improved Relation Network[J]. Acta Optica Sinica, 2020, 40(8): 0811005 Copy Citation Text show less
    Overall architecture of infrared aircraft classification learning model with small samples
    Fig. 1. Overall architecture of infrared aircraft classification learning model with small samples
    Architecture of module. (a) Embedding module; (b) relation module
    Fig. 2. Architecture of module. (a) Embedding module; (b) relation module
    Illustration of few-shot learning datasets under meta learning training mode
    Fig. 3. Illustration of few-shot learning datasets under meta learning training mode
    Pseudo-code for learning algorithm
    Fig. 4. Pseudo-code for learning algorithm
    Partial examples of three datasets. (a) mini-ImageNet dataset; (b) Infra-object dataset; (c) Infra-aircraft dataset
    Fig. 5. Partial examples of three datasets. (a) mini-ImageNet dataset; (b) Infra-object dataset; (c) Infra-aircraft dataset
    Test accuracy and loss curves on mini-ImageNet dataset. (a) 5-way 1-shot; (b) 5-way 5-shot
    Fig. 6. Test accuracy and loss curves on mini-ImageNet dataset. (a) 5-way 1-shot; (b) 5-way 5-shot
    Relationship between test accuracy and adequacy of training samples on mini-ImageNet dataset
    Fig. 7. Relationship between test accuracy and adequacy of training samples on mini-ImageNet dataset
    Accuracy comparison of ground to air infrared aircraft classification
    Fig. 8. Accuracy comparison of ground to air infrared aircraft classification
    Convolution kernel size5-way 1-shot5-way 5-shot
    3×381.2391.28
    5×581.4190.92
    7×779.3387.07
    9×975.9882.30
    Table 1. Accuracy of convolution kernel size estimation%
    ModelFine-tune5-way 1-shot5-way 5-shot
    Baseline-nearest-neighbor[19]N41.08±0.7051.04±0.65
    Baseline-linear[20]Y42.11±0.7162.53±0.69
    Meta-learner LSTM[19]N43.44±0.7760.60±0.71
    MAML[21]Y48.70±1.8463.11±0.92
    Matching network[13]Y42.4058.00
    Prototypical network[14]F49.42±0.7868.20±0.66
    RelationNet[15]F50.44±0.8265.32±0.70
    Improved relation networkF54.89±1.0269.87±0.75
    Table 2. Accuracy of each model on mini-ImageNet dataset%
    GroupTraining datasetTest dataset5-way 1-shot5-way 5-shot8-way 1-shot8-way 5-shot
    1Infra-object+Infra-aircraftInfra-object+Infra-aircraft86.25±1.2594.84±0.6677.82±1.7991.11±0.63
    2Infra-objectInfra-aircraft84.37±1.3193.66±0.7677.56±1.4690.58±0.64
    3mini-ImageNetInfra-aircraft78.92±2.7890.76±1.2474.44±3.2886.34±1.95
    4Infra-object+mini-ImageNetInfra-aircraft82.79±0.7594.51±0.8278.47±0.9489.92±1.02
    Table 3. Accuracy of model for infrared aircraft classification on different training datasets%
    Lu Jin, Shijian Liu, Xiao Wang, Fanming Li. Infrared Aircraft Classification Method with Small Samples Based on Improved Relation Network[J]. Acta Optica Sinica, 2020, 40(8): 0811005
    Download Citation