Digital image forgery (DIF) is a prevalent issue in the modern age, where malicious actors manipulate images for various purposes, including deception and misinformation. Detecting such forgeries is a critical task fo...Digital image forgery (DIF) is a prevalent issue in the modern age, where malicious actors manipulate images for various purposes, including deception and misinformation. Detecting such forgeries is a critical task for maintaining the integrity of digital content. This thesis explores the use of Modified Error Level Analysis (ELA) in combination with a Convolutional Neural Network (CNN), as well as, Feedforward Neural Network (FNN) model to detect digital image forgeries. Additionally, incorporation of Explainable Artificial Intelligence (XAI) to this research provided insights into the process of decision-making by the models. The study trains and tests the models on the CASIA2 dataset, emphasizing both authentic and forged images. The CNN model is trained and evaluated, and Explainable AI (SHapley Additive exPlanation— SHAP) is incorporated to explain the model’s predictions. Similarly, the FNN model is trained and evaluated, and XAI (SHAP) is incorporated to explain the model’s predictions. The results obtained from the analysis reveals that the proposed approach using CNN model is most effective in detecting image forgeries and provides valuable explanations for decision interpretability.展开更多
文摘目的研究CASIA2优化的NK和KS公式预测ICL(V4c)拱高与实际拱高的一致性。设计诊断技术评价。研究对象合肥爱尔眼科医院行ICL植入手术患者68例(68眼)。方法应用Nomogram在线公式计算68眼ICL型号,根据在线公式推荐结果分为12.1 mm、12.6 mm、13.2 mm ICL三组,再采用CASIA2测量以上三种型号组术后1天实际拱高(V1d12.1、V1d12.6、V1d13.2)、术后1个月实际拱高(V1m12.1、V1m12.6、V1m13.2),分别计算与CASIA2术前应用NK、KS公式预测拱高值的差异性、相关性和一致性,以及CASIA2内置的NK、KS公式术前推荐的ICL(V4c)型号与在线公式推荐的ICL(V4c)型号之间的符合率。主要指标实际拱高与公式预测拱高的相关性和一致性。结果全部术眼(68眼)术后1天实际拱高为(528.8±176.9)μm,术后1个月有所降低,为(476.8±161.2)μm。NK公式预测拱高为(656.2±204.5)μm,KS公式为(707.1±173.8)μm。Nomogram在线公式与NK公式之间总符合率为55.88%,与KS公式总符合率为42.65%。3种型号ICL(V4c)组术后1天实际拱高(V1d12.1、V1d12.6、V1d13.2)和1个月的实际拱高(V1m12.1、V1m12.6、V1m13.2)分别与NK公式预测拱高(VNK)、KS公式预测拱高(VKS)之间比较,除了V1d12.1组与VKS组拱高差异无统计学意义(t=-2.056,P=0.079)外,其余各型号组实际拱高均低于NK公式和KS公式预测的拱高(P均<0.05)。NK公式、KS公式的预测拱高与V1d均呈线性关系(Y=0.4728*X+218.6,Y=0.5450*X+143.4);与V1m也均呈线性关系(Y=0.3985*X+215.3,Y=0.4604*X+151.2)。68例患者术后V1d与NK公式、KS公式预测的拱高之间组内相关系数(ICC)分别为0.541、0.536,克隆巴赫系数分别为0.702、0.700;术后V1m与NK公式、KS公式预测拱高之间ICC分别为0.492、0.495,克隆巴赫系数分别为0.659、0.662。Bland-Altman分析V1d与NK公式、KS公式预测的拱高之间95%LoA分别为-486.5~231.7μm、-509.6~153.0μm,V1m与NK公式、KS公式预测的拱高之间95%LoA分别为-543.4~184.5μm、-560.6~99.86μm。结论CASIA2优化的NK和KS公式预测的ICL(V4c)拱高均较实际拱高偏高,两种公式对小型号组ICL(V4c)预测的准确性稍好。
文摘Digital image forgery (DIF) is a prevalent issue in the modern age, where malicious actors manipulate images for various purposes, including deception and misinformation. Detecting such forgeries is a critical task for maintaining the integrity of digital content. This thesis explores the use of Modified Error Level Analysis (ELA) in combination with a Convolutional Neural Network (CNN), as well as, Feedforward Neural Network (FNN) model to detect digital image forgeries. Additionally, incorporation of Explainable Artificial Intelligence (XAI) to this research provided insights into the process of decision-making by the models. The study trains and tests the models on the CASIA2 dataset, emphasizing both authentic and forged images. The CNN model is trained and evaluated, and Explainable AI (SHapley Additive exPlanation— SHAP) is incorporated to explain the model’s predictions. Similarly, the FNN model is trained and evaluated, and XAI (SHAP) is incorporated to explain the model’s predictions. The results obtained from the analysis reveals that the proposed approach using CNN model is most effective in detecting image forgeries and provides valuable explanations for decision interpretability.
