Short-read accuracy is important for downstream analyses such as genome assembly and hybrid long-read correction. Despite much work on short-read correction, present-day correctors either do not scale well on large data sets or consider reads as mere suites of k-mers, without taking into account their full-length read information. We propose a new method to correct short reads using de Bruijn graphs, and implement it as a tool called Bcool. As a first st ep, Bcool constructs a compacted de Bruijn graph from the reads. This graph is filtered on the basis of k-mer abundance then of unitig abundance, thereby removing from most sequencing errors. The cleaned graph is then used as a reference on which the reads are mapped to correct them. We show that this approach yields more accurate reads than k-mer-spectrum correctors while being scalable to human-size genomic datasets and beyond. The implementation is open source and available at http: //github.com/Malfoy/BCOOL under the Affero GPL license.