The classic visual navigation system of mobile robots relies on manually designed feature extraction algorithms, which leads to insufficient recognition accuracy in autonomous navigation. The research proposes an improved convolutional neural network (I-CNN) method to optimize the path recognition status of the visual navigation system of mobile robots. The experimental results show that the recognition accuracy of I-CNN reaches 92.3% after 1000 iterations, which is significantly higher than 87.5% of the traditional CNN. The model demonstrated a faster convergence speed and lower loss value during the training process. Its accuracy improved to 95.5% after 1000 iterations, significantly outperforming the unoptimized version. I-CNN can converge rapidly and improve the recognition accuracy in the path recognition task.