Research on Anoectochilus Roxburghii Unknown Category Detection Module Based on Confidence Correction

Different strains of Anoectochilus roxburghii have different medicinal values. For seaking high profit, there are also many adulterated Anoectochilus roxburghii sold on the market. In order to establish the quality standard of Anoectochilus roxburghii, it is very meaningful to classify different strains of Anoectochilus roxburghii and its counterfeits. Classification models based on deep neural network (DNN) output high confidence label predictions, and can obtain high classification accuracy for samples of given categories. However, when they are used for out-of-distribution samples, i.e. samples of unknown categories, the DNN models often produce false classification. In order to solve this problem, an output confidence correction strategy is proposed. By using the input reconstruction error of the autoencoder, the high corrections for unknown categories can be reduced. Experimental results for identifying five strains of Anoectochilus roxburghii and two counterfeits show that the proposed detection module has good detection effect. Meanwhile, the proposed module does not need to modify the original classification model at the bottom, and can be easily transplanted to other classification models using softmax probability value as output confidence.