The present invention relates to technology capable of effectively classifying objects from compressed images such as H.264 AVC, H.265 HEVC and the like in general. More specifically, for example, unlike existing technology in which objects are recognized and classified through complex image processing in regard to a compressed image generated by a CCTV camera, syntax information (e.g., a motion vector, a coding type) obtained by parsing compressed image data is used to extract an area in the image, in which a certain meaningful motion exists, namely a moving object area, and then, an image of the moving object area is regarded as an object candidate area to be inputted into a convolution neural network (CNN) to obtain an object classification result. In particular, since motion vector patterns of the moving object area are regarded as a training data group to perform machine learning for a deep neural network, a motion vector RPN (MRPN) with enhanced localization performance is formed, and then, the MRPN is applied to a forward pass of the CNN to preprocess the image of the moving object area. Thus, if one single moving object area includes a plurality of objects, the objects are separated to obtain an object classification result for each of the objects.;COPYRIGHT KIPO 2020
展开▼