Wireless optical channel key extraction serves as an effective method for achieving physical layer security， wherein legitimate parties generate a key sequence by detecting and estimating the features of the optical channel. Eavesdroppers， however， may capture key-related information through optical signals outside the legitimate party's receiving aperture. To address this challenge， a novel key extraction scheme was proposed. Initially， an enhanced generative adversarial network model was utilized to enable legitimate parties to accurately estimate channel features relevant for key extraction from the measured channel sequences. Subsequently， a random quantization alternating method was employed on the estimated channel feature sequences to derive the initial keys independently. Experimental results demonstrate that the key sequences generated by the legitimate parties exhibit high consistency， with the correlation coefficient between the channel feature sequences estimated by Alice and Bob reaching 0.998 3 in a 25 dB SNR environment. Furthermore， the initial key disagreement rate produced by Alice and Bob is 1.3×10^-4 ，which is significantly reduced by the proposed scheme. The shared key obtained after information negotiation by the legitimate parties successfully passes the NIST randomness test， underscoring the efficacy and robustness of the proposed method in enhancing key extraction reliability.