A Combined Deep Learning and Semi-supervised Classification Algorithm for LS Area

Abstract

In real world, there are many areas with Large images but only Small labelled (we called LS area), in there supervised and unsupervised algorithm can’t work well, but semi-supervised technology exploiting patterns both in labelled and unlabeled data to get labels can work well. The classification accuracy directly depends on the features extracted from the images. Recently, with the emergence and successful deployment of deep learning techniques for image classification, more research on getting features is directed to deep learning techniques. This paper proposes a combined semi-supervised classifier and pre-trained deep CNN model algorithm—CDLSSC (Combined Deep Learning and Semi-Supervised Classification) for LS area. The transfer learning that has been tested and verified in some areas is used to extract features in this algorithm. The method CDLSSC is evaluated on three image datasets and achieves superior performance. We apply it to the Terra-Cotta Warriors image classification area and get super results, which means that it can be used in cultural relic’s area successfully.