This paper proposes a novel loss function for training a distributed convolutional neural network (DisCNN) to recognize only a specific positive class. By mapping positive samples to a compact set in high-dimensional space and negative samples to Origin, the DisCNN extracts only the features of the positive class. An experiment is given to prove this. Thus, the features of the positive class are disentangled from those of the negative classes. The model has a lightweight architecture because only a few positive-class features need to be extracted. The model demonstrates excellent generalization on the test data and remains effective even for unseen classes. Finally, using DisCNN, object detection of positive samples embedded in a large and complex background is straightforward.
Figures & tables
DisCNN-8
DisCNN-1
VGG
input(96x96RGB)
input(96x96RGB)
input(96x96RGB)
Conv3-64
Conv3-64
Conv3-64
BN
BN
BN
ReLU
ReLU
ReLU
Max pooling
Max pooling
Max pooling
Conv3-32
Conv3-32
Conv3-128
Table 1: Model Architectures. The convolutional layer parameters are denoted as “conv¡receptive field size¿-¡number of channels¿”. The batch normalization are denoeted as BN.
Network
DisCNN-8
DisCNN-1
VGG
Number of parameters
0.149
0.028
3.096
Table 2: Number of parameters (in millions).
Positive
Negative
True
TP = 787
FN = 175
False
FP = 13
TN = 1425
Table 3: Confusion matrix
threshold
Precision
Recall
F1
0
Positive
0.985
0.816
0.892
Negative
0.889
0.992
0.937
1
Positive
0.968
0.924
0.945
Negative
0.960
0.983
0.972
2
Positive
0.941
0.964
0.953
Negative
0.983
0.971
0.977
Table 4: Evaluation Metrics of different thresholds
threshold
Precision
Recall
F1
0
Positive
0.965
0.778
0.862
Negative
0.863
0.980
0.918
1
Positive
0.931
0.894
0.912
Negative
0.945
0.965
0.955
2
Positive
0.886
0.944
0.914
Negative
0.974
0.945
0.959
Table 5: Evaluation Metrics of different thresholds for unseen classes