Salient target detection method for 360-degree panoramic image

The invention discloses a salient target detection method for a 360-degree panoramic image. According to the method, an equidistant columnar projection picture is input, and six equi-angular cube chartlets can be obtained after the equidistant columnar projection picture enters a model and is preprocessed. Then, a depth feature extraction network extracts multi-level depth space features of the two pictures respectively; feature enhancement is performed on the dual-type feature fusion network by adopting a channel attention mechanism, and depth features of each layer of the equidistant columnar projection and the equal-anglar cube chartlets are fused by utilizing a multi-scale cavity convolution module; and finally, multilayer composite feature information obtained by fusion is input into a multilayer information fusion memory network in sequence, high-level and low-level semantic information is memorized, screened, and fused step by step, and finally a saliency target prediction map is obtained. According to