Butterfly image retrieval is a challenging issue requiring the feature representation not only to be sensitive to the subtle inter-class difference but also to remain robust to large intra-class variations. Fan-beam projection is a mathematical tool originally applied to computed tomographic (CT) reconstruction of objects. In this paper, we introduce it for the first time into object recognition field. Separated fan-beam projection followed by Gaussian convolutions of different widths are designed to extract multiscale invariant features, patch-projection angles (PPA) and texture-projection angles (TPA), for separately depicting the patch patterns and texture properties of butterfly images. The PPA and TPA are then treated as heterogeneous co-occurrence patterns to be fused by a 2D histograms as final feature representation. We present a comprehensive experimental evaluation including image retrieval at species and subspecies levels, complementarity to deep-learning features, invariance and robustness. All the results consistently show the superior performance of the proposed method over the state-of-the arts.