A convolutional neural network (CNN) for fractional interpolation of inter prediction is proposed because of the poor interpolation effect of traditional interpolation filters and the deep learning methods, which only generate half pixel samples, or need to train the corresponding model for each pixel position and quantization parameter (QP), or introduce additional information as input. Based on the dense residual network, the model combines multi-scale distortion feature extraction structure and sub-pixel convolution to increase the accuracy of feature extraction and generate fractional pixels. The characteristics of fractional interpolation task are analyzed and the data set with true distortion is constructed. The model directly generates fractional pixel samples and can adapt to arbitrary quantization parameters (QP). Experimental results verify the efficiency of the method. Compared with H.265/HEVC, this method achieves 2% in bit saving on average under low-delay P configuration. Compared with similar methods, the overall performance has also been improved.