Curvelet decomposition is a multiscale analysis method defined for 2D and 3D signals that can represent curve-like features with great sparsity. A genuine method based on histograms of curvelets is proposed for content-based texture image retrieval. The accuracy of the method is analyzed for rotation invariance, curvelet scale-orientation size, and bin size. The results are given with precision-recall graphs. Experimental results on the Brodatz database show promising results for the proposed method compared to curvelet subband statistical features.