Structure restoration plays an important role in image completion. Though much progress has been made, existing techniques are ineffective to restore plausible complex structures, due to the high complexity of structure registration between neighboring regions. We address this challenge by taking two measures. The first is to get the images whose structures are potential to be well merged with the structures around the hole. This can be achieved by effectively retrieving the mass images on the internet, which provide enough candidates. The second is to coherently blend the retrieved structures with the structures around the hole, which is by suppressing the connection differences between these structures. Experimental results show that we can easily complete images, especially with complex structures or for filling large holes.