Many multimedia analysis and processing applications can benefit from better image understanding. However, accurate understanding of what the underlying visual content represents still remains a very challenging problem. Systems need to look beyond the pure matrix of intensity values and exploit other types of information. Actually, humans do not solve visual problems based just on the captured pixel data. Many non-visual clues and diverse types of external information are exploited, including prior knowledge, prior experience, contextual and collaborative information. Intelligent systems can often incorporate them to simplify the problem, and can also integrate other types of signals, such depth and infrarred images. Often this additional information can be critical to better address the problem or increase the quality of the result. This special issue contains 10 papers, selected from the inital 25 submissions, after two rounds of blind review.

Priveleged or expert information can help to guide learning by exploiting indirect cues. In “Facial Expression Recognition through Modeling Age-related Spatial Patterns” [6], Wang et al describe a system that leverages age information, available only during training, to improve facial expression recognition. Age-related spatial expression patterns carry crucial information that has not been exploited previously. In particular, the authors implement this system using Bayesian networks modeling facial geometric features, outperforming existing approaches. The paper “Maximum Margin Hashing with Supervised Information” [8], by Yang et al, describes a large margin approach to learn hashing functions in a supervised way. Leveraging label information helps to obtain hashing functions that better represent the semantic similarity between images. The paper “Discriminative Sparse Representation for Face Recognition” [10], by Zhang et al, proposes to use facial structure information to increase face recognition performance. The weights are learned for different face locations based on information entropy, providing an effective prior and structure information. A sparse coding formulation combined with this weighting scheme results in more discriminative representations. Experiments on face recognition benchmarks show good results. In “High-Order Graph Matching Kernel for Early Carcinoma EUS Image Classification” [9], Zhang et al address the problem of early carcinoma detection in electronic ultrasonography images. The method is based on graph matching, and the authors propose a new kernel that preserves better the underlying topological structure of the graph. The paper “Discriminative Sparse Neighbor Coding” [1], by Bai et al, proposes a new coding method addressing some shortcomings of traditional sparse coding strategies. They include a module that selects discriminative features for each class, and then another module that discards non-informative visual words. Finally they obtain class-specific low dimensional subspaces, with results achieving state-of-the-art performance.

Combining multiple heterogeneous signals or multiple features extracted from the same signal can also outperform methods using a single one. In “Features combination for art authentication studies: brushstroke and materials analysis of Amadeo de Souza-Cardoso” [4], Montagner et al tackle the interesting problem of art authentication, in the particular case of the modernist painter Amadeo de Souza-Cardoso. In addition to conventional RGB analysis, which is used to analyze brushstroke characteristics, their system incorporates hyperspectral imaging and elemental analysis to determine the pigments present in the painting and compare them with those used by the painter. “Camouflage Performance Analysis and Evaluation Framework Based on Features Fusion” [7], by Xue et al, proposes a model to evaluate the quality of camouflage features. Their model uses nonlinear fusion to combine multiple image features that measure the degree to which the target and surrounding background differ with respect to background-related and internal features. Subjective experiments show that scores predicted by the system correlate with the difficulty human assessors had to detect camouflaged targets. In “Pointwise and Pairwise Clothing Annotation: Combining Features from Social Media” [5], Nogueira et al exploit social media data to automatically annotate clothes, including the different garment items. The authors formulate the task as a multi-label and multi-modal classification problem, proposing two approaches (pointwise and pairwise). Their experiments show a significant improvement compared with related methods.

External information can be used in image enhancement and other image processing operations. The paper “Image Super-Resolution based on Multi-kernel Regression”, by Li [2] et al, describes a data-driven method to obtain high resolution images from blurred ones using multi-kernel regression, and an efficient algorithm to implement it. The key information to obtain the details that enhance the images is learned from an external dataset. In “LWT- QR decomposition-based robust and efficient image watermarking scheme using Lagrangian SVR” [3], Mehta et al address the problem of watermaking images, that is, embedding an image with an external signal in a such a way that the signal cannot be perceived, but recovered if necessary. Their method uses the coefficients from a lifting transform followed by the QR decomposition to train a support vector regression model. The model is used to embed the watermark and then to extract it when necessary.

These ten papers cover a wide range of image analysis and processing applications and scenarios in which additional information is beneficial. We hope the readers will find interesting ideas in them. Finally, to conclude, the guest editors would like to express their gratitude to the authors who submitted manuscripts and to the reviewers who generously contributed reviewing them.