Aiming at the lack of robustness of traditional methods due to uneven illumination, equipment vibration and background noise in the detection of rail surface defects in complex industrial scenarios, this paper proposes a hierarchical detection method based on dynamic low-rank projection with eigenvalue constraints. The method alleviates feature mixing by introducing an eigenvalue constraint to force the separation of high-frequency defect features and low-frequency background features; then constructs an adaptive low-rank projection matrix and optimizes the projection direction using backpropagation to reduce the dependence on manual parameter adjustment; and finally, designs a hierarchical optimization strategy combining the low-rank constraint and the sparse error term to suppress the unstructured noise and accurately locate the sparse defects. Experimental evaluations on the RSDD dataset demonstrate that the proposed method achieves 97% recognition accuracy on both Type-I and Type-II data subsets, surpassing existing methods by a significant margin. By adjusting the low-rank parameter 1 ≤ r ≤ 4, the method can adapt itself to different data complexity, and the minimum detection size is 0.5 × 0.2 mm, which meets the demand of industrial refinement detection.