While an individual video frame may exhibit low resolution, multiple images of the same subject are typically collected. This motivates the use of super-resolution techniques - that is, using computational methods to synthesize a single higher-resolution image from multiple lower-resolution images of the subject. A number of authors have investigated the effects of resolution and super-resolution in facial recognition, and the results have indicated both a high degree of dependence between recognition performance and resolution, and significant potential for performance improvement through super-resolution. Despite the promise of super-resolution, the literature had lacked a comprehensive and quantitative presentation of the improvement that may be realized on realistic data. Here Hu and coworkers present such an analysis. They thoroughly describe the collection of representative test data and evaluate the improvements given using previously-published super-resolution and facial recognition algorithms. A range of relevant test scenarios are considered (e.g., varying ranges and numbers of images), with many cases exhibiting significant improvements with super-resolution. Given the thoroughness of this work, the improvements realized and the quantitative nature of the presented results, I expect this manuscript to be an important touchstone for future work on improved facial recognition through super-resolution.
You must log in to add comments.