Abstract: In order to solve some practical engineering problems in coal mining and processing,such as the height ad- justment of shearer’s drum and preliminary gangue discharge in coal preparation plants,an effective approach to coal- rock recognition via statistical modeling in dual-tree complex wavelet domain was proposed. Firstly,coal or rock images were decomposed using multi-level dual-tree complex wavelet transform ( DTCWT). Secondly,a strategy for rotation- invariance enhancement was presented. According to this strategy,high-frequency sub-bands generated from every level DTCWT were sorted in descending order by the product of mean and variance of sub-band coefficient modulus. Third- ly,an assumption that generalized gamma distribution could be regarded as the underlying distribution of high-frequen- cy sub-band coefficient modulus was made,and the scale-independent shape estimation ( SISE) equations based pa- rameter estimation method was employed to determine the parameters of generalized gamma distribution. Finally,the automatic discrimination between coal and rock images was completed using similarity measurement with respect to rel- ative entropy. Experimental results demonstrate that the generalized gamma distribution statistical model has strong power to distinguish between coal and rock images. To some extent,the strategy presented for rotation-invariance en- hancement helps to increase the correct recognition rate of coal-rock recognition,and it is much more flexible to make a trade-off between correct recognition rate and time complexity by means of this strategy. The proposed approach outperforms several other existing ones in terms of correct recognition rate,the time complexity of which is still acceptable. This study could provide some guidance and reference for future practice on unmanned coal mining.