This paper presents a 3D deconvolution method for fluorescence microscopy that reached the first place at the ''the 3D Deconvolution Microscopy Challenge'' held during ISBI 2013. It uses sparse coding algorithm to learn 2D ''high resolution'' features that will be used as a prior to enhance the resolution along depth axis. This is a three steps method: (i) deconvolution step with total variation regularization, (ii) denoising of the deconvolved image using learned sparse coding, (iii) deconvolution using denoised image as quadratic prior. Its effectiveness is illustrated on both synthetic and real data.