Dataset for image classification with knowledge

Deep learning applied to raw data has demonstrated outstanding image classification performance, mainly when abundant data is available. However, performance significantly degrades when a substantial volume of data is unavailable. Furthermore, in situations where distinguishing between distinct classes is challenging, such as in fine-grained image classification, deep architectures struggle to achieve satisfactory performance levels.
Utilizing a priori knowledge alongside raw data can enhance image classification in demanding scenarios. Nevertheless, only a limited number of image classification datasets given with a priori knowledge are currently available, thereby restricting research efforts in this field. This paper introduces innovative datasets for the classification problem that integrate a priori knowledge. These datasets are built from existing data typically employed for multilabel multiclass classification or object detection. Frequent closed itemset mining is used to build classes and their corresponding attributes (e.g. presence of an object in an image) and then to extract a priori knowledge expressed by rules on these attributes. The algorithm for generating rules is described.