Early power modeling and analysis using electronic system-level methodology enables designers to explore energy saving opportunities more efficiently at a higher abstraction level. However, power modeling for third party IPs are challenging due to the limited observability and unknown architecture details. To model the data dependency for blackbox IPs, several works rely on adopting Hamming distance of input data to approximate the switching activity, which might be not enough for modeling complex IPs such as image signal processors (ISP). This work introduces a content-aware line-based power modeling method for ISP by training an associated energy table. To effectively estimate ISP energy consumption which involves many two-dimensional data processing, this work presents a direct energy-mapping strategy using pixel luminance and gradient. Moreover, an iterative box-constrained least-squares estimation and its associated constraint refinement scheme is proposed to increase the robustness of the trained energy table even with limited training data. Simulation results show that the proposed method can reduce at least 11.54% of average error and 55.52% of max error as compared to the existing content-blind power model.