Intelligent robotic systems process images captured by camera devices to get high-level semantic concepts. Current image captioning approaches use convolutional neural networks as encoder and recurrent neural networks as decoder. Various modalities are treated differently and parallel computation is not fully allowed due to the chain architecture. To address these issues, we present a parallel hierarchical neural network based on encoder-decoder architecture to generate descriptions for images. Only convolutional neural networks are adopted here, and the hierarchical framework takes less steps to capture distant dependencies. Masked positional self-attention mechanism is utilized in the decoder to improve the performance. The fixed sized windows make it possible for parallel computation. This generative model uses unified architecture for different modalities, reaching a BLEU-1 score (the higher the better) over 0.7 with a higher training speed.