Abstract: This paper proposes a grouping method of convolution structure for feature extraction of images. The model increases the depth of network, and the network parameters changes smaller. Compared with the traditional CNN, the network using the module of this paper increases the recognition rate from 50.1% to 52.2% on the data set Caltech256. On the data set 101_food, the recognition accuracy increased from 66.3% to 68.9%. Experimental results show that the network structure in the paper improves the network performance.