Abstract—Intelligent reflecting surface (IRS) is a developing technology that can significantly enhance the efficiency of wireless communications. It achieves this by smartly adjusting the signal reflections at several passive reflecting elements. The channel estimation is a crucial problem in implementing a viable IRS-assisted communication system. Deep learning (DL) has attracted considerable attention for tackling physical layer design issues because of its capacity to acquire intricate patterns from data with less computing complexity and enhanced resilience. This paper proposes a channel estimation model with a correlated channel based on DL that employs the Minimum Mean-Squared Error (MMSE) criterion. Specifically, we design and train the Convolutional Neural Network (CNN) architecture using received signals to simultaneously estimate both the direct channel between the transmitter and receiver, as well as the cascaded channel that incorporates the IRS’s reflection. Furthermore, the numerical results demonstrate that incorporating IRS and the proposed DL-based channel estimation technique leads to substantial performance gains over conventional channel estimation methods. Specifically, at low SNRs (-5 dB), the DL-based approach exhibits NMSE values of approximately 0.7659, whereas at higher SNRs (25 dB), NMSE values decrease to around 0.0022. These findings underscore the efficacy of the proposed solution in mitigating the adverse effects of channel impairments.