Small Target Image Caption Generation Based On Feature Redirection

Image caption is automatically converting an image into a text description that contains scenes, objects, actions, and other key information. Because it is necessary to understand the content of the image and output a sentence with correct semantics, this task covers the fields of computer vision and natural language processing. When generating image captions, due to problems such as the small size of small targets, small proportion of the entire image, and lack of available feature information, existing image caption algorithms have flaws in extracting features of small targets, resulting in insufficient existing image caption algorithms. In order to meet the feature extraction of small targets in ordinary scenes. This article proposes a feature redirection model. Firstly, different levels of dilated convolution are used to capture multi-scale features and reduce small-size target features. Finally, the feature values of different scales are fused to obtain richer feature information and generate richer image caption. We conducted extensive experiments on the MSCOCO datasets and verified the superiority of the model.