To overcome the problems of low efficiency in manual meter reading and inaccurate recognition for double half-characters by existing image recognition methods, an image-based intelligent meter reading system based on narrow band Internet of things (NB-IoT) and lightweight convolutional neural network (CNN) is designed. Firstly, the image acquisition terminal uses NB-IoT module to upload the meter image collected by the camera to cloud platform. Then, the method of local feature extraction and matching is applied to estimate an affine transform matrix and convert the input meter image to the coordinate space of the template image, and every sub-image of reading character is segment out. Finally, a multi-label classification-based lightweight CNN model is proposed to recognize these sub-images, and the final reading result is obtained by post-processing. Experimental results indicate that the sleep current of the image acquisition terminal of the designed system is less than 10 μA,which can ensure two Li/SOCl₂ batteries working for more than 5 years, and that the proposed CNN model based on multi-label classification can accurately recognize both single characters and double half-characters and has achieved a character accuracy rate of 96.36% and a reading accuracy rate of 94.15%, which is superior to other recognition algorithms.