Solution for Authenticity Identification of Typical Target Remote Sensing Images

TL;DR

This paper introduces a weakly supervised RGB model for high-precision authenticity identification of remote sensing images, utilizing pseudo-labels and a two-step mask generation and model fine-tuning process.

Contribution

It presents a novel two-stage approach combining pseudo-mask generation with model fine-tuning, improving accuracy in remote sensing image authenticity detection.

Findings

Achieved a test score of 90.7702.

Effectively handled imprecise mask generation.

Enhanced model accuracy by selective feature use.

Abstract

In this paper, we propose a basic RGB single-mode model based on weakly supervised training under pseudo labels, which performs high-precision authenticity identification under multi-scene typical target remote sensing images. Due to the imprecision of Mask generation, we divide the task into two sub-tasks: generating pseudo-mask and fine-tuning model based on generated Masks. In generating pseudo masks, we use MM-Fusion as the base model to generate masks for large objects such as planes and ships. By manually calibrating the Mask of a small object such as a car, a highly accurate pseudo-mask is obtained. For the task of fine-tuning models based on generating masks, we use the WSCL model as the base model. It is worth noting that due to the difference between the generated pseudo-Masks and the real Masks, we discard the image feature extractors such as SRM and Noiseprint++ in WSCL, and select the unscaled original image for training alone, which greatly ensures the match between the image and the original label. The final trained model achieved a score of 90.7702 on the test set.