A Framework for Image Dark Data Assessment.

Image dark data, whose content and value are not clear, consistently occupy the storage space but hardly produce great value. Blindly applying data mining techniques on these data is highly likely to bring disappointed result and waste large resource. Therefore, it is of great significance to assess the dark data before data mining to help the user cognize the data. However, there are several challenges in dark data assessment work. First, the similarity between images must be objectively measured under aunified standard to help the user understand the evaluation values of dark data. Second, it is important to capture semantic features with generalization ability. Third, it is challenging to design an efficient assessment scheme to support large-scale datasets. To overcome these challenges, we propose an assessment framework which includes offline calculation and online assessment. In offline calculation, we first transform unlabeled images into hash codes by our developed Deep Self-taught Hashing (DSTH) algorithm which can extract semantic features with generalization ability, then construct a semantic graph using restricted Hamming distance, and finally use our designed Semantic Hash Ranking (SHR) algorithm to calculate the overall importance score (rank) for each node (image), which takes both the number of connected links and the weight on edges into consideration. During online assessment, we first translate the user’s query (semantic images) into hash codes using DSTH model, then match the data contained in the dark data via a predefined Hamming distance query range, and finally return the weighted average value of these matched data to help the user cognize the dark data. The results on real-world dataset show our framework can apply to large-scale datasets, help users evaluate the dark data by different requirements, and assist the user to conduct subsequent data mining work.