Positioning algorithms based on visible light communication can help solve the problem of insufficient positioning accuracy in some special occasions such as indoors, basements, and underground. However, the accuracy of existing visible light positioning algorithms is difficult to improve further, and most of them remain in the simulation stage without experimental verification. In the previous visible light positioning algorithms based on deep learning, such algorithms can be divided into positioning algorithms based on received light signal intensity and positioning algorithms based on light source images according to the source of data. The positioning algorithm based on received light signal intensity receives the light signal intensity from each LED light source in turn, and obtains the positioning result based on the characteristics of the light signal intensity. Similarly, the positioning algorithm based on light source image receives the LED light source image and obtains the positioning result by analyzing the characteristics of the light source. This article proposes a positioning algorithm (MIF-VLP) based on the attention mechanism to fuse light Signal Intensity Information (RSS) and image information. The MIF-VLP algorithm uses ResNet-18 as the backbone network of the image, and maps RSS into a vector through word embedding, and then adjusts the output of ResNet-18 to make them have the same dimension. The fusion method of the attention mechanism is based on the image, so the input of the attention layer is multiple vectors, and the output is only one vector. The advantages of the algorithm are that, firstly, the algorithm uses both the light signal intensity information and the image information, which makes up for the overfitting problem caused by the single data, improves the generalization ability of the model and the final positioning accuracy. Secondly, the algorithm performs a permutation and combination preprocessing on the received light signal intensity, treats the input light signal intensity as a sequence, ignores the order and number of light signal intensities from multiple LEDs, and reduces the dependence on the environment to a certain extent. The traditional positioning algorithm based on the intensity of received light signals has strict requirements on the input order and number of data, because each light signal intensity represents the characteristics of an LED. In addition, the MIF-VLP algorithm can expand the size of the data set under the same experimental environment, so that the model can converge in fewer epochs and improve the generalization performance of the model. In comparison, the article selected the RSS-BP algorithm based only on RSS data and the CNN algorithm based only on light source images in the experimental stage. The CNN algorithm also uses ResNet-18 as the backbone network. The experimental results show that in the experimental environment of 2 m×2 m×1.8 m, the average positioning error of the MIF-VLP algorithm reaches 5 mm, which is 80.7% higher than that of the RSS-BP algorithm based on RSS information and 87.5% higher than that of the convolutional neural network algorithm based on image information; the maximum positioning error of the MIF-VLP algorithm reaches 8.9 cm, which is 41.4% lower than that of the RSS-BP algorithm based on RSS information and 19.1% lower than that of the convolutional neural network algorithm based on image information. The minimum positioning error of the MIF-VLP algorithm reaches 0.5 mm, which is much lower than the 1cm positioning error of the RSS-BP algorithm and the 2 mm positioning error of the CNN algorithm. Overall, among all the positioning points, the MIF-VLP algorithm has only one point with an error greater than 2 cm, which shows the stability of the MIF-VLP algorithm. Analysis shows that the reason for the large error at this point may be due to reflection or measurement error, because this point is located at the edge and is easily affected by reflected light.