An image forming apparatus, such as a laser printer, ink jet printer, or a thermal transfer, which comprises a neural network. The apparatus improves image quality by reducing a zigzag included, e.g., in input image data, reducing the circuit size by reducing the number of bits having small weights for an input combination, and concurrently correcting the size of center, left and right picture elements in a predetermined window. The neural network outputs correction data for the size and position of the center dot in a window in response to an input of dot image data in a window or subdot pattern exposure data for the center dot. Additionally, the neural network uses any of the three values, +1, -1 and 0, for the coefficient of input combination for a hidden layer neuron, e.g., after a teacher pattern is learned. Further, the neural network outputs correction data for 3.times.N subblocks of picture elements obtained by dividing blocks of picture elements in the center, left and right of a window.