Abstract: The rapid calculation of blast pile fragment size distribution has been a focal point in both academia and industry due to its significant application on optimizing blasting effects and reducing mining costs. In this study, the high-resolution orthophoto datasets of open-pit mine blast piles were acquired using nap-of-the-object photogrammetry techniques, and a deep learning algorithm for fragment size distribution recognition was proposed to assess blasting effectiveness and optimize mining costs. To enhance the feature extraction of different rock fragmentation sizes, we introduced a switchable atrous convolution module and a recursive feature pyramid refinement module. Fourier descriptors were utilized to establish statistical distributions of the blast piles, while the cumulative passing volume curve was employed in place of the cumulative passing rate. The results demonstrated the effectiveness of the proposed algorithm: the mean fine fragmentation rate on the surface of the target blast pile was 8.90 %, and the mean large block rate on the surface was 4.69 %. The high fine fragmentation rate and low large block rate indicate that the blasting parameters can be further optimized, and the cost can be reduced.