To address the contradiction between watermarking invisibility and robustness,a zero-watermarking method based on dense connection and redundant feature elimination network (DCRFENet) is proposed.Firstly,in order to resist different image attacks,the dense connection module is designed to extract shallow and deep layer image robust features from different convolution layers.Meanwhile,to enhance the uniqueness of zero-watermarking,the redundant feature elimination module is presented to eliminate redundant information and enhance image valid features by learning weight of inter-feature and intra-feature.Secondly,valid features and robust features are fused to generate the final image feature map,which is used for robustness training.Finally,based on training of DCRFENet,the image feature map is divided into blocks,and zero watermark is obtained by comparing each feature value within the block with its average value.The experimental results show that the average bit error rate (BER) is lower than 0.03 for CIFAR10,COCO and VOC datasets. Moreover,compared with the existing methods,the proposed zero-watermarking model is robuster to trained attacks,untrained attacks and hybrid attacks.