Redshift is one of the basic parameters of galaxies. The number of photometric images is enormous relative to the spectrum. A large number of known galaxies have only photometric images and no spectra. It is, therefore, common to obtain redshift values from photometric images rather than spectra. This paper constructs the Galactic Redshift Regression Network (GRRnet), a convolutional neural network for estimating photometric redshift from galaxy images. It has a deeper network layer than previous similar methods and adds an attention mechanism to help it focus on more useful information. Based on GRRnet, this paper further proposes a two-step strategy, GRRnet-C-R: the first step is to classify the galaxies according to the redshift roughly; the second step is to perform regression estimation according to the classified categories, and finally merge them. This strategy can significantly reduce the error of photometric redshift estimation. The data in this paper are all from the Sloan Digital Sky SurveyDR16. The relevant data of each galaxy includes the composite images of the three bands of g, r, and z, the photometric values of the five bands of u, g, r, i, and z, and the observed Spectral redshift for labeling. In the pre-processing process, the photometric image is cut to the size of 50×50, to ensure that most of the galaxies can be framed while the computation is reduced. Since the input size of the comparison algorithm NetZ is 64×64, to keep the input size consistent, use the function cv2. resize to change the image size to 64×64. The experimental results show that the mean squared error (MSE) of GRRnet-C-R reaches 0.001 46, which is 22.3%, 21.9% and 18.0% lower than that of random forest (RF), eXtreme Gradient Boosting (XGBoost) and NetZ, respectively. The linear regression coefficient of determination of GRRnet-C-R reached 0.948, which achieved a good model fitting effect. Experimental results show that this two-part strategy can effectively reduce the error of metering redshift estimation, which provides a new idea and method for subsequent metering redshift estimation.