In this paper a new approach for blurred image restoration is presented. Our algorithm is based on human vision which zooms back and forth in the image in order to identify global structures or details. Deconvolution parameters are estimated by an edge detection and correspond to the ones of a chosen edge detection model. The segmentation is obtained by merging multiscale information provided by multiscale edge detection. The edge detection is achieved by using a derivative approach following a generalization of Canny-Deriche filtering. This multiscale analysis performs an efficient edge detection in noisy blurred images. The merging leads to the best local representation of edge information across scales. The algorithm deals with a mixed (coarse-to-fine/fine-to-coarse) approach and searches for candidate edge points through the scales. Edge characteristics are estimated by the merging algorithm for the chosen model. Scale, direction and amplitude informations allow a local deconvolution of the original image. The noise problem is not considered in this work since it does not disturb the process. Results show that this method allows non-uniformly blurred image restoration. An implementation of the whole algorithm in an intelligent camera (DSP) has been performed.