An Efficient Novel Approach for Detection of Recolored Image Using Deep Learning for Identifying the Original Images in Public Surveillance

Abstract

Background: Pitie et al. used the adjustment in themes of a picture to align the desired classification of another pixel by a common activity in image editing. Whereby, they are difficult to find the right mapping because it needs multiple rigorous methods such as matching, combining the channels, or evolving the contrast. Image recoloring is a process of changing the color of the image and is one of the most significant purposes of manipulation. Recoloring is a process of adjusting the colors of the image by changing one color to another or by adjusting color intensity with another color, adjusting the brightness or contrast, and converting colors to shades of gray. Purposes: Image coloring identification technique is derived from the deep learning techniques that includes CNN architecture design by using different layers. This also includes the actual image and two channels relying on inter-channel correlation and illumination mapping to find out the recolor pictures. Based on the usage of the layers, the accuracy and sensing of time may vary. A new architecture based on the stated layers and the use of a filter is designed to increase accuracy, decrease sensing time, and also reduce misclassification when compared. This architecture is different from the existing architecture for the detection of recolored images. Methods: This study is designed with a new architecture shown in Fig. 1 to improve the accuracy and reduce the time taken while training the network. To solve the problem, Convolutional Neural Networks is one of the techniques used, and Deep Neural Networks classes that rely on Multi-Layer Perception and Back Propagation which is mainly used for image processing, classification, segmentation, and also for other auto-related data and differs from conventional. Results: The output for an image is shown in table whether it is a recolored image or a normal image only after training the network model. Based on the training of the network, the accuracy is considered as mini-batch accuracy where 100% accuracy is acquired. Conclusions: Both inter-channel and illumination consistency are employed to help feature extraction. Datasets are applied and used different filters size and the different number of filters. A major effect on classification accuracy is predicted by designing a structure of filters. Basic recommendations for future work are using more intricate and deeper models for unpredictable problems. The model is more effective by the integration of DNN with the theory of enhanced learning.