Near real time enhancement of geospatial imagery via systolic implementation of neural network-adapted convex regularization techniques Y. Shkvarko a , A. Castillo Atoche a,b,⇑ , D. Torres-Roman a a Department of Telecommunications, CINVESTAV (Centro de Investigaciones y Estudios Avanzados del IPN), Unidad Guadalajara, Mexico b Department of Mechatronics, Autonomous University of Yucatan, Mexico article info Article history: Available online 12 June 2011 Keywords: Remote sensing imagery Network of Systolic Arrays Neural networks Regularization abstract In this paper, we address a new approach for near-real-time enhancement of large-scale Geospatial and aerial remote sensing (RS) imagery that aggregates descriptive and Bayesian convex regularization par- adigms for solving the image reconstruction inverse problems with efficient systolic-based neural net- work (NN) computing. This task is approached via Hardware–Software (HW/SW) codesign oriented at the Field Programmable Gate Array (FPGA) digital implementation that unifies the NN-adapted image enhancement/reconstruction techniques with a novel efficient computational architecture based on a Network of Systolic Arrays (NSA). We demonstrate how such unification reduces drastically the compu- tational load of the real-world RS image enhancement/reconstruction tasks resulting in efficient numer- ical algorithms suitable for quasi real-time NN-adapted implementation with the existing generation of the FPGA-based digital processors that implement the proposed NSA computational architecture. Ó 2011 Elsevier B.V. All rights reserved. 1. Introduction Understanding of the Geospatial remote sensing (RS) imagery, as well as solution of diverse pattern recognition tasks depend drastically on the provided RS image quality (Henderson and Lewis, 1998; Barrett and Myers, 2004; Shkvarko et al., 2001; Shkvarko, 2002, 2010). That is why, enhancement of the noise cor- rupted low resolution RS imagery before its intelligent analysis is an important information processing task that could substantially improve the decision making results. The problems of enhance- ment of the large-scale RS imagery are computationally extremely expensive and are not suitable for implementation with the exist- ing digital signal processors (DSP) if run the image enhancement/ reconstruction procedures in the original concurrent algorithmic format (Henderson and Lewis, 1998; Kung, 1988). In this paper, we address a new approach for (near) real-time enhancement of the large-scale Geospatial and aerial RS imagery that aggregates descriptive and statistical regularization paradigms for solving the image reconstruction inverse problems adapted for efficient systolic-based neural network (NN) computing. Within this con- text we unify the recently developed descriptive experiment de- sign regularization (DEDR) and the fused Bayesian regularization (FBR) nonparametric paradigms for high and super-high resolution image enhancement, reconstruction and fusion. Next, the near real time implementation task is approached via Hardware–Software (HW/SW) codesign oriented at the Field Programmable Gate Array (FPGA) digital implementation that unifies the NN-adapted enhancement/reconstruction techniques with a novel efficient computational architecture based on a Network of Systolic Arrays (NSA) (Kung, 1988). We demonstrate how such unification reduces drastically the computational load of the real-world RS image enhancement/reconstruction tasks resulting in efficient numerical algorithms suitable for quasi real-time NN-adapted implementa- tion with the existing generation of the FPGA-based digital proces- sors that implement the proposed NSA computational architecture. 2. Problem formulation Consider an RS imaging experiment performed with coherent array imaging radar/synthetic aperture radar (SAR) systems (Hen- derson and Lewis, 1998; Barrett and Myers, 2004; Shkvarko et al., 2001; Shkvarko, 2002, 2010). Following (Henderson and Lewis, 1998; Shkvarko, 2002, 2010), we define the DSP-oriented discret- ized model of the M  1 observation data vector u by specifying the stochastic equation of observation (EO) in a vector–matrix form u ¼ Se þ n; ð1Þ where e defines the K  1 vector of a discrete-form representation of the scene complex reflectivity function (Henderson and Lewis, 1998), S is the M  K (M P K) signal formation operator (SFO) 0167-8655/$ - see front matter Ó 2011 Elsevier B.V. All rights reserved. doi:10.1016/j.patrec.2011.05.018 ⇑ Corresponding author at: Department of Telecommunications, CINVESTAV (Centro de Investigaciones y Estudios Avanzados del IPN), Unidad Guadalajara, Mexico. E-mail addresses: shkvarko@gdl.cinvestav.mx (Y. Shkvarko), acastillo@gdl.cin- vestav.mx, acastill@uady.mx (A. Castillo Atoche), dtorres@gdl.cinvestav.mx (D. Torres-Roman). Pattern Recognition Letters 32 (2011) 2197–2205 Contents lists available at ScienceDirect Pattern Recognition Letters journal homepage: www.elsevier.com/locate/patrec