The present invention provides a deep learning-based image recognition method for a CCTV. The deep learning-based image recognition method for a CCTV of an image recognition system having a plurality of CCTVs and a control center server comprises the steps of: (a) allowing an image recording unit to photograph and record an object to be controlled by using a built-in camera and to output recording data on shape information and moving information; (b) allowing a background extracting unit to receive the recording data, separate and extract an image of the object to be controlled from a background image within a current frame through the shape of the photographed object to be controlled and to output the extracted data; (c) allowing an image analyzing unit to receive the extracted data, analyze a differential image between the current frame and a next frame, extract a spatial feature of the object to be controlled, and generate a vector space to output context data, feature data, and face data; and (d) allowing a feature vector calculating unit to receive the context data, the feature data, and the face data and to compare the received data with a prestored pattern by using a deep learning technique to calculate a feature vector of the object to be controlled. The feature vector calculated from each of the plurality of CCTVs is transmitted to another CCTV connected in a machine-to-machine manner, thereby integrally tracking and monitoring the object to be controlled through an intelligent connection. According to the present invention, reliability of a product and data processing speed are improved.
展开▼