A fully convolutional dual-path neural network model with improved cross-entropy loss function is proposed to solve the problem of identifying partial discharge maps of electrical equipment. Using the partial discharge map as the model input, the deep and shallow features of the map are extracted by two channels using different size convolution kernels, and then performing feature fusion. The convolutional layer is used instead of the fully connection layer to preserve more spatial correlation between PD features. The improved cross-entropy loss function can make the model more suitable for the situation of imbalanced datasets. The experimental results show that the accuracy of the improved FCN dual-path feature fusion recognition method reaches 99.31%, which can accurately identify the partial discharge map, and the amount of model parameters is smaller.