There are a large number of analogy meters due to complex electromagnetic environment in transformer substation and these meters need manual reading, which makes it difficult to automated manage transformer substation. Currently, most meter automatic reading methods rely on preacquired high quality image, in which meter targets are big in size and locate in the middle and surface of meter is parallel with the camera. This needs lots of prior meter measurement and camera calibration, which fails to meet the requirements of actual use in transformer substation. In order to solve the problem mentioned, this paper presents a complete meter detection and recognition method. First, meter location within current visual field is obtained through a convolutional neural network model. Then the difference between target center location and camera visual field center location, as well as size percentage of target, are calculated. Camera state, including camera location and camera scaling factor, is adjusted according to the calculation result. After that, high quality image of meter target is acquired through perspective transform, which eliminates the image distortion caused by nonparallelism between meter and camera. Finally, locations of dial and pointer of the meter are obtained by conducting Hough Transform to the meter image, and meter reading is achieved. Results of actual experiments with transformer substation indicate that, maximum of reading error is as low as 1.82%. The proposed method can obtain accurate and stable performance with multiple kinds of meters in complicated background, which meets the demand of practical application in transformer substation.