Sketches contain only simple lines and contours, which have completely different characteristics from natural images with rich colors and details. However, the current neural networks are mostly designed for natural images and cannot adapt to the sparseness of sketches. Aiming at this problem, this study proposes a sketch-based image retrieval method based on deformable convolution. First, the Berkeley edge detection algorithm is used to transform the natural image into edge map to eliminate domain differences. Then replace part of the standard convolution in the convolutional neural networks with deformable convolution, so that the network can fully focus on the outlines of the sketches. Finally, sketches and edge maps are sent to the network separately, and extract the fully connected layer features as feature descriptors for retrieval. Experimental results on the benchmark dataset Flickr15k show that the proposed method can effectively improve the accuracy of sketch-based image retrieval compared with existing methods.