We have proposed charge-domain time-compressive LiDAR image sensors that compress the scene response using an indirect time-of-flight (ToF) sensor and compressive sensing. The scene’s impulse response is reconstructed in post-processing. Because these image sensors are based on an indirect ToF sensor using multitap charge modulators, they are suitable for high spatial resolution. They can achieve high distance measurement accuracy by reconstructing the scene response to estimate ToF. Since the ToF estimation relies on the reconstructed scene response, its fidelity is crucial. In this paper, we compared a deep learning model inspired by an iterative optimization method with a conventional total variation-based iterative method. We assumed an application to autonomous robotic vacuum cleaners in which multipath interference between the floor and the wall may introduce significant depth errors. In both simulations and experiments, we confirmed that regularization based on a U-Net in ADMM-Net achieves higher fidelity in the reconstructed waveform than TVAL3.