The internal damage of concrete has obvious characteristics of discreteness and randomness, and the detection of internal damage characteristics is an important content of concrete mesoscopic research. To solve the problem of low precision of existing models, this paper proposes a double-head Cascade R-CNN model with feature sharing to detect the damage features of concrete CT images. Firstly, a Cascade R-CNN network model, which combines spatially sensitive fc-head (fully connected head) and spatially correlation conv-head (convolution head), is constructed to effectively identify the spatial information of the damage feature. Then, the classification information of each level of the detection network is merged through the feature sharing method, which improves the precision of low IOU(intersection over union) threshold (0.5sim0.7) ROI (regions of interest) detection tasks. The results show that the average precision of the proposed method is 91.31%, which is 3.04% higher than that of the original. The average accuracy of low IOU threshold (0.5sim0.7) ROI is improved by 1.49%. The proposed model can better detect the mesoscopic damage part from the concrete CT image with the features of high precision, simple computation and easy engineering realization.