A Multi-layer Feature Parallel Processing Method for Image Captioning

Most image captioning methods based on neural networks use high-level features extracted by CNNs, but it is difficult for high-level features to retain the information of small objects, so the generated description cannot meet more fine-grained requirements. To solve the above problems, we propose a multi-layer feature parallel processing method for image captioning, which feeds each layer of features to each stacked layer of the decoder in a certain order, thereby using multi-feature expression to generate a more fine-grained description. We provide two design schemes for the proposed multi-layer feature parallel processing method: Sequential Parallel Connection(SPC) and Reverse Parallel Connection(RPC). This work focuses on exploring a more effective and robust model connection method that can generate finer-grained descriptions. Extensive experiments in the COCO dataset show that our connection method can generate better quality sentences.