To remove impulse noise from image before some other process, an adaptive switching filter based-on neural network noise detector is proposed. This method describes each pixel with pixel value and its neighborhood characteristics and takes these as inputs of the neural network to identify pixels which are likely to be contaminated by noise with the trained neural network automatically. According to the idea of switching filter, the noisy pixels detected are processed by mean filter with adaptive window size, and only noise-free pixels of the window are involved in the average computation. Compared with some other common filters, the experimental result shows that this BP neural network has high accuracy of salt and pepper noise detection. Besides, this filtering process is superior in denoising effect, details preserving and time consuming reduction without manual intervention.