https://hal.univ-brest.fr/hal-03221196Rondel, NadineNadineRondelThomson Broadband SystemsBurel, GillesGillesBurelThomson Broadband SystemsCooperation of multi-layer perceptrons for the estimation of skew angle in text document imagesHAL CCSD1995Skew Angle Estimation (SAE)Hand-written textsAngles of Arrival (AOA)Array-processingCooperation of Neural NetworksMaximum Likelihood[SPI.SIGNAL] Engineering Sciences [physics]/Signal and Image processingBurel, Gilles2021-05-07 19:22:432023-03-18 03:10:372021-05-07 19:22:43enConference papers10.1109/ICDAR.1995.6021221Estimating the skew angle in text document images can be a crucial problem in optical character recognition. Based on a new sensor array processing technique, an original solution to skew angle estimation (SAE) is proposed. Thanks to the reformulation of the SAE problem in the framework of angle of arrival theory, a fast and accurate method is presented that is based on the cooperation of two neural networks. The first neural net is a three-layer perceptron receiving on input the values of the correlation matrix of the signals; the output is a "rough" estimation of the angle to estimate. This gross estimate is then used to initialize the weights of a second multi-layer perceptron (MLP). The second MLP is built in order to perform a maximum likelihood-like optimization, therefore reaching good performances. The system, though trained on simulated radar data, shows good performances on noisy handwritten texts.