This paper presents a robust and efficient Convolutional Neural Network (CNN)-based channel estimator for fifth-generation (5G) Orthogonal Frequency Division Multiplexing (OFDM) systems. While conventional Least Squares (LS) and Minimum Mean Square Error (MMSE) estimators degrade significantly in high-mobility, non-linear, and frequency-selective fading environments, the proposed framework treats the time-frequency resource grid as a spatial image, enabling implicit learning of complex fading dynamics without explicit statistical modeling. The model is trained on 105 synthetic channel realizations spanning Rayleigh, Rician (K = 5 dB), AWGN, CDL-A, CDL-B, and TDL-A channel profiles and validated through rigorous MATLAB simulations. Key quantitative results demonstrate: (i) a 7-fold BER reduction over MMSE at 20 dB SNR on CDL- A (1.0 × 10−3 vs. 7.0 × 10−3); (ii) a 3–5 dB NMSE improvement across the full 0–30 dB SNR range; (iii) robust performance under Doppler spreads up to 300 km/h with less than 0.5 dB BER penalty; (iv) a 50% reduction in pilot overhead while maintaining superior MSE performance; and (v) spectral efficiency within 0.35 bits/s/Hz of the perfect-CSI Shannon bound. With a measured inference latency of 0.8 ms and a lightweight design of 2.3 × 106 parameters, the proposed CNN-CE is validated as a practically deployable and resource-efficient technology for 5G and beyond-5G (B5G) networks.