Many researches are conducted to improve Hopfield Neural Network performance especially for speed and memory capacity in different approaches. However, there is still a significant scope for developing HNN using Optical Logic Gates. We propose a new model of HNN based on all-optical XNOR logic gates for real time image recognition. Firstly, we improved HNN toward optimum learning and converging operations. We considered each unipolar image as a set of small blocks of 3-pixels as vectors for HNN. In addition, the weight matrices which have stability of unity at the diagonal perform clear converging in comparison with no self-connecting architecture. Synchronously, matrix-matrix multiplication operation would run optically in the second part, since we propose an array of all-optical XOR gates, which uses Mach-Zehnder Interferometer for neurons setup. The controlling system is to distribute and invert signals to achieve XNOR function. The preliminary experiment show positive results of the proposed system.