Remote sensing images are an important basis for humans to obtain information on the surface. However, due to the limitations of sensor industrial technology and the influence of the sensor's working environment, remote sensing images generally contain fringe noise, which seriously damages image information, and cannot be used directly. Different from many current advanced convolutional network image stripe noise removal methods which focus on optimizing the network structure, the method proposed in this paper mainly splits the input-output data of the convolutional network. Through the reorganization process, the lightweight convolutional network with fewer layers and fewer channels has a good image stripe denoising effect. However, such models using convolutional networks are only suitable for stripes with a certain width, and stripes with inappropriate widths will seriously affect the processing effect of this type of model. If strips in large-size images such as remote sensing images are removed, and a stripe is hundreds of pixels wide, the stripe removal effect of this type of model will be very weak. Therefore, this paper proposes a technique of splitting-reorganizing the image of the input-output convolutional network, so that the above-mentioned models can have an excellent removal effect on stripes of any width.