In order to solve the problem that the deep learning network model has limited ability to recognize image features in low resolution, which affects the recognition accuracy and speed, a convolutional neural network-conjugate gradient (CNN-CG) model is proposed. By improving the conjugate gradient map feature recognition algorithm, the convolutional neural network model is built to recognize and predict the image features. The input image is projected from three dimensions to 32, 64, 128 and 256 dimensions through convolution layer and activation function in turn. The encoder outputs high-dimensional classification features, and the image category information is obtained after full connection layer. Compared with other models on the same data set, the proposed model performs well in terms of recognition accuracy, convergence and image data analysis. The recognition accuracy of the training set is 100%, and the accuracy of the test set is 80.26%, which is better than that of the general model. The results show that the model has a good effect on the traffic sign image recognition application, and shows strong robustness and ablation, which is suitable for the optimization algorithm of the traffic sign image recognition model in the automatic driving application scene, and can be extended to the feature recognition of images in other fields.