Binary classification study in the weakly supervised barcodes detection

The barcodes detection based on the construction of a class activation map uses several neural network models of binary classification. In addition to the classical architecture of a neural network with one neuron and the Sigmoid activation function, two neurons and the Softmax activation function can be used. In this paper, we investigate the influence of this factor on the quality of barcodes detection using weakly labeled data. The best search quality is obtained using two neurons in a fully connected layer of the binary classification model, viz. 0.725 precision, 0.674 recall, 0.698 F1, the quality using one neuron, viz. 0.574 precision, 0.573 recall, 0.573 F1.