SALIENT OBJECT DETECTION METHOD AND SYSTEM FOR WEAK SUPERVISION-BASED SPATIO-TEMPORAL CASCADE NEURAL NETWORK

摘要

Provided is a salient object detection method for use in the field of video and image recognition, wherein a spatio-temporal cascade neural network comprises a first full convolutional network and a second full convolutional network; the method comprises: inputting a current frame image of a video to be detected into the first full convolutional network to obtain a spatial prior image (S1); generating a temporal prior image according the current frame image and an optical flow image thereof (S2); performing an element operation on the spatial prior image and the temporal prior image to obtain a spatio-temporal prior image (S3); and inputting the spatio-temporal prior image and the next frame image into the second full convolutional network to obtain a spatio-temporal salient image (S4). When detecting a salient object of a video which has a complex scene, spatial prior information of a video frame image and optical flow-based time prior information are integrated, thus achieving the elimination of static salient regions and the generation of the final spatio-temporal salient image within a dynamic scene, such that more abundant information may be acquired within the dynamic scene, thus improving accuracy and robustness.