The invention discloses a convolutional-neural-network-based vehicle model refined classification system. A vehicle image super-amount classification method is provided to solve problems of unsharp features and difficult feature extraction and brand and type information of a vehicle is obtained based on an image analysis, so that license plate and vehicle logo information that is easy to shield or tamper does not need to be used. The system especially can be applied to fields of traffic control, traffic inspection, traffic statistics, and criminal investigation of fake-licensed cars, license-free cars, and license-shielded cars. With the method, a problem of super-amount classification for vehicle brands and types with images can be solved. A hybrid convolutional neural network having a plurality of sub networks is used; a local picture and a global picture of a vehicle are inputted into all sub networks simultaneously, wherein a vehicle face picture, a vehicle body picture, and a vehicle tail picture are inputted into different sub networks specifically; and scoring is carried out by using a sub network grading layer, so that a classification result is obtained.