With the continuous development of financial market, the traditional option pricing model faces many limitations when dealing with complex options. With its powerful nonlinear fitting ability, neural networks can effectively capture the complex relationship between option prices and many factors. In this paper, by optimizing the network structure, the number of network layers is reduced to three, and sparse connection, convolution layer and regularization technology are introduced, which significantly reduces the number of parameters and improves the calculation efficiency. At the same time, the training speed of neural networks is further accelerated by using a parallel computing framework, PyTorch, and GPU hardware acceleration. In the aspect of improving accuracy, this paper determines the best learning rate, iteration times and hidden layer dimension through hyperparameter optimization, and adopts L1 regularization and Dropout technology to prevent over-fitting. The empirical analysis shows that the optimized neural network performs well in option pricing, its training time is greatly shortened, its prediction accuracy is significantly improved, its model complexity is reduced, and its generalization ability is enhanced. The research results provide an efficient and accurate new method for pricing complex options, which has important theoretical and practical significance.