셀프 어텐션과 합성곱 신경망을 이용한 JPEG 이중 압축 탐지

초록

In this paper, we propose a network for double JPEG detection. We generated a new dataset using the quantization tables obtained in [3]. The proposed network consists of 1-dimensional convolutional layers and a transformer encoder[6]. We extract DCT coefficient histogram and quantization table from previously mentioned dataset, and utilize them as input to the network. Also, we localize the forgery region?single JPEG block?by detecting if the JPEG block has been compressed more than once or not. This proposed approach achieved higher accuracy than baseline [3] in detecting single JPEG.