Feedforward neural network for temporal speckle pattern optimization

MARCELO GUZMÁN; GUSTAVO MESCHINO; ISABEL PASSONI; MARCELO TRIVI; HECTOR RABAL

Santiago de Compostela

Congreso; 23rd Congress of the International Commission for Optics; 2014

International Commission for Optics

Speckle interferometry techniques is a useful tool for characterization of dynamic processes based on the analysis of the temporal history of the speckle pattern (THSP) [1]. They have been used in multiple biological and industrial applications [2]. However, obtaining THSP´s involve the acquisition of many large images, typically 300x300 or 512x512 pixels2. This paper proposes the use of a feedforward neural network [3] which is trained with the resulting values of the 2D wavelet decomposition of reduced THSP´s (20x20, 20x50, 20x100 and 100x100 pixels2). Speckle images used in this case are obtained from the well know study of the paint drying process. The drying time has been previously set by gravimetric techniques. Several settings for neural networks are tested and the best one is chosen when achieving the lowest error. Thus, the size of the reduced THSP and network with the lowest error is selected. Finally the selected samples with new network is evaluated.