Memory positional encoding for image captioning

Abstract

Transformer-based architectures represent the state-of-the-art in image captioning. Due to its natural parallel internal structure, it cannot be aware of the order of inputting tokens, so the positional encoding becomes an indispensable component of Transformer-based models. However, most of the existing absolute positional encodings (APE) have certain limitations for image captioning. Their spatial positional features are predefined and cannot been well generalized to other forms of data, such as visual data. Meanwhile, each positional features are decoupled from each other and lack internal correlation, therefore which affects the accuracy of spatial position context representation of visual or text semantic to a certain extent. Therefore, we propose a memory positional encoding (MPE), which has generalization ability that can be applied to both the visual encoder and the sequence decoder of the image captioning models. In MPE, each positional feature is recursively generated by the learnable network with memory function, making the current generated positional features effectively inherit the genetic information of the previous $n$ positions. In addition, existing positional encodings provide positional features with fixed value and scale, that means, they provide the same positional encoding for different inputs, which is unreasonable. Thus, to address the previous issues of scale and value of current positional encoding methods in practical applications, we further explore dynamic memory positional encoding (DMPE) based on MPE. DMPE dynamically adjusts and generates positional features based on different input to provide them with unique positional representation. Extensive experiments on the MSCOCO validate the effectiveness of MPE and DMPE.