SOft Computing applications for Complex EnviRonments

Research Lines

  • Forensic identification:
  • We are a pioneering research group applying soft computing and computer vision techniques to the forensic identification field since we started our collaboration with the Physical Anthropology Lab of the University of Granada headed by the prestigious Dr. Miguel Botella in 2006. After almost ten years we have published in the most prestigious journals in both the forensic and artificial intelligence fields. We have coordinated several research projects, highlighting the FP7 European Project “MEPROCS: New methodologies and protocols in forensic identification by craniofacial superimposition”. We are currently collaborating with the most prestigious forensic labs in Europe such as CAHID, LABANOF, or the Israel National Police. Our main contribution to the field refers to the automation of the Craniofacial Superimposition (CFS) technique. CFS is intended to determine whether a skull found corresponds to a particular person through the superimposition of the former over one or more photographs of the face of the latter and the posterior morphological analysis of their correspondence. An hybridization of computer vision, fuzzy logic and evolutionary algorithms resulted in the most advanced computer-aided system to assist the forensic anthropologist in the identification of a missing person by CFS, Face2Skull. This system has been developed in collaboration with the ECSC after obtaining the corresponding international patent. Further information about this research line can be found at SOCOVIFI webpage. We are now exploring new applications to other skeleton-based profiling and identification techniques.

  • Economic modeling:
  • The IRAM lab is involved in a technological alliance with a marketing consultancy called R0D Brand Consultants. Our developments include the design of two different kinds of economic models. The first one is aimed for strategic branding. It is based on a macro-level, top-down approach implemented by a systems dynamics methodology which incorporates fuzzy logic to model relations between tangible and intangible macroscopic variables (e.g. brand equity, loyalty, HR budget, etc.), machine learning and social network analysis techniques to select the most relevant variables, and genetic algorithms for strategic action optimization. The model allows clients to incorporate both the existing data in their companies (reputation studies, surveys, etc.) along with their expert knowledge. It has already become a commercial product called Zio® SD.

    The second model deals with consumer mix modelling. It thus follows a micro-level, bottom-up approach considering an agent-based computational economics model to recreate competitive markets for modelling and predicting sales, touchpoints, perceptions, awareness, and word-of-mouth. Simulated consumers have an active decision making procedure based on their perceptions and awareness of the existing brands in the artificial market and they are also connected by a social network modeling diffusion processes. The system, called Zio® ABM, allows for market understanding and for the handling of what if scenarios to observe the impact of investments on the consumers in terms of sales evolution.

  • Assembly line balancing:
  • We have used evolutionary multi-objective optimization and multi-objective ant colony optimization algorithms to solve a challenging real-world problem, the time and space assembly line problem (TSALBP), which involves to achieve optimal assignments of a subset of tasks to each station of the assembly line of a plant with respect to two or three conflicting objectives to be minimized: its cycle time, its number of stations and their area. This framework emerged thanks to the observation of a real automotive industry plant belonging to Nissan and located in Barcelona, Spain, as this research was performed in collaboration with the Prothius Industrial Engineering Chair (formerly Nissan Endowed Chair) of the Technical University of Catalonia, headed by Prof. Joaquín Bautista.

  • Medical image segmentation and registration:
  • We coordinated the FP7 European Project "MIBISOC: Medical Imaging using Bio-inspired and Soft Computing" that supported our work on medical image segmentation and registration. Image segmentation is commonly defined as the partitioning of an image into non-overlapping regions that are homogeneous with respect to some visual feature, such as color or texture. In many medical imaging applications, segmentation algorithms play a crucial role by automatically identifying anatomical structures and other regions of interest. Such algorithms are nowadays in the core of multiple tasks, like quantification and measurement of tissue volumes, localization of pathologies or computer-integrated surgery. It is important to highlight that manual segmentation is not only tedious and time consuming but, sometimes, also inaccurate, hence the importance of developing automatic and accurate segmentation methods. We have applied Soft Computing and Machine Learning in image segmentation using deformable models. In its most general formulation, image registration (IR) is the task of aligning two or more images in order to establish a spatial correspondence of their common content. Such images usually have the same or a similar subject but have been acquired under different conditions, such as time and viewpoint, or by multiple sensors. In medical image analysis, IR is a key technology that allows to "fuse" visual information from different sources. Applications include combining images of the same subject from different modalities, detecting changes before/after treatment, aligning temporal sequences of images to compensate for motion between scans, image guidance during interventions and aligning images from multiple subjects in cohort studies. We have tackled medical IR with different metaheuristics following feature-based and intensity-based approaches.

  • Visual science map modeling and mining:
  • Visual science maps (scientograms) are a powerful tool to represent and analyze scientific information. They comprise high quality, schematic visualizations for huge scientific domains. The Scimago research group has developed a methodology to design scientograms based on the combined use of Thomson’s Web of Science JCR® or Elsevier Scopus® co-citation information and social network analysis through the use of the Pathfinder network pruning algorithm. This methodology is integrated in their SCImago Journal & Country Rank web information system. The SOCCER lab members have developed new network pruning algorithms for scientogram design that are incorporated into this system. These methods significantly reduce the Pathfinder run time, allowing us to generate scientograms of very large scientific domains (even of the whole World production) in an on-line fashion.
    Meanwhile, we have also made significant developments to the analysis of scientograms by means of single and multi-objective graph mining methods. Information science experts are not able to deal with the complex task of analyzing a large number of scientograms. We have thus performed scientogram mining in order to analyze and compare the structure of scientific fields and research fronts in maps of the same (taken at different periods of time) or different domains (looking for similarities between different countries scientific productions). In particular, we have developed pioneering, general-purpose graph mining techniques using multi-objective and multimodal evolutionary algorithms to evaluate subgraphs based on different conflicting criteria such as support and size. Our methods can be customized to different learning tasks and are able to retrieve a Pareto set of non-dominated, meaningful subgraphs from a structural database in a single run. In addition, this approach allows us to uncover less frequent subgraphs that describe the underlying phenomena from different angles, revealing novel information that otherwise would have been hidden by uninformative frequent descriptions.