Conditional Information-Based Adaptive Fusion Local Redrawing Model with Data Extension Framework

Due to the huge differences in scene complexity and object scale of different images, local image redrawing is a very challenging task. Existing work generally focuses on the completeness and uniformity of objects before and after local redrawing, however, it neglects the reasonableness of the object scale and the naturalness of the object scene articulation. To address this problem, this paper proposes a novel framework for image local redrawing, called Adaptive Fusion Local Redrawing Model (AFLRM), which is capable of adaptively generating accurate masks and optimizing mask edge processing. Specifically, we implement an adaptive mask generation technique that can dynamically adjust the mask size according to the complex scale relationships between scenes and objects adaptively. At the same time, the fusion prompt guides the dynamic adjustment of mask edge blurring and the number of reserved pixels, which effectively reduces the transition traces between the generated content and the original image, and realizes a more natural fusion redrawing effect. And based on this, a consistent framework for automatically expanding image data is designed, which can effectively solve the problem of insufficient experimental data in some fields through the automatic redrawing of data. Extensive experiments on the public datasets COCO and EditBench show that our AFLRM achieves high-quality local redrawing, is general and robust, and provides a good balance between uniformity before and after object redrawing as well as object scale articulation rationality.