The image splicing localization algorithm based on dual-stream Faster R-CNN achieves a good performance because it considers both the color image and its corresponding noise image as inputs. However, it still has the following two drawbacks, it only achieves block-level precision and the noise images generated by SRM filter are likely to contain a lot of redundant non-forged semantic features. Therefore, this paper proposes a pixel-level image splicing localization model based on dual-stream Faster R-CNN. Regarding the first drawback, a fully convolutional neural network branch is added for pixel-level localization. Regarding the second one, the steganalysis rich model is replaced by error level analysis noise model for noise map extraction. Experimental results show that the proposed algorithm improves the accuracy by nearly 10% compared with some existing algorithms.