Due to the influence of lighting and harsh weather, the traditional image processing methods have low recognition efficiency in visual navigation of inspection robots. This paper proposes a visual navigation method for power inspection robots based on image preprocessing and semantic segmentation. An image enhancement method based on the adaptive gamma correction method is proposed to solve the influence of strong light, weak light and uneven illumination on the image. Aiming to the exposure conditions, the LSTM prediction model is used to adaptively adjust the camera angle to eliminate the exposure and improve the good exposure of the image. The improved DenseNet is used to semantically segment the navigation path and extract the path target area, fitting the robot's forward route through the pixel value distribution of the target area and calculate the offset, which provides the key parameters of robots to adjust the driving posture. Template matching is used to determine the direction, location and bifurcation signs in the navigation path. Experimental results show that the algorithm could effectively solve the problem of low recognition accuracy caused by lighting and adverse weather, and improve the accuracy of navigation and positioning of inspection robots in complex environments.