Studies on Performance of Image Splicing Techniques Using Learned Self-Consistency

Abstract

Abstract In recent years, improper image manipulation has become a significant issue across many sectors, such as in the education field, politics, entertainment sector, and social media platforms. In various society sectors, images play a crucial role in interpreting the facts and number of fraud incidents. The fraud and fact can be detected by image manipulation, which is gaining more and more attention in today’s world. Detecting such manipulations images is a major challenging problem concerning many fragmented images’ unavailability as training data. A learning algorithm has been proposed for recognizing many visually spliced images and manipulations in the image. The proposed algorithm also utilizes the existing recorded image Exchangeable Image File Format (EXIF) metadata as a supervisory image for training the model to identify the image is self-consistent, i.e., whether the content of the photo has been generated using a single imaging pipeline. The proposed algorithm is implemented to the errand of identifying and localizing the image splices. An excellent performance improvement has been observed on several benchmarks, despite not using any spliced image data while training. This model will stand out as a vital source to detect morphed images in various industries.KeywordsSelf-ConsistentExchangeable image file formatImaging pipeline