Modeling Automated Image Watermarking Using Meta-heuristic-based Deep Learning with Wavelet Approach

Accessing digital media has become quite a simple result of the fast rise of multimedia in the case of network technology. As a result, safeguarding intellectual property necessitates a greater focus on image watermarking. Distinct image watermarking systems have been introduced for this purpose; however, they have limitations with transparency and robustness. In the sphere of digital watermarking, multimedia copyright protection plays a critical part. The practice of extracting and embedding a watermark discreetly on a carrier image is called digital image watermarking. Digital watermarking is successful in securing digital data; it also has sparked a lot of study attention nowadays. Deep learning networks combined with wavelet-oriented approaches for image watermarking have gotten a lot of interest these days. Conventional watermarking techniques, on the other hand, cannot provide blindness, resilience, and automated extraction and embedding all at the same time. In this circumstance, this paper motivates to offer an improved approach for generating watermarked images with elevated invisibility using deep learning with a novel wavelet-based technique. Initially, after gathering the data, image griding is performed to partition the images into grids, thus making the image suitable for efficient feature extraction. Then, the two techniques named Deep feature extraction by Convolutional Neural Network and Neighboring-based features are extracted. Using these features, the Modified Deep Neural Network (MDNN) is used for choosing the regions for embedding the watermark. Here, the training algorithm of DNN is optimized by the Squirrel Search Algorithm (SSA) and Grey Wolf Optimization (GWO), known as Squirrel Search-Grey Wolf Optimization (SS-GWO). Once the regions are selected, watermark embedding is performed by the Adaptive Discrete Wavelet Transform (ADWT) with filter coefficient optimization by the same SS-GWO based on a newly derived fitness function. Accordingly, the message extraction is achieved using the same ADWT with the embedding key. Throughout the results, the mean of SS-GWO-MDNN + ADWT is 31.25%, 10.53%, 16.67%, and 23.53% improved than SSA-MDNN + ADWT, GWO-MDNN + ADWT, PSO-MDNN + ADWT, and JA-MDNN + ADWT regarding Gaussian filtering attack for dataset 3. The simulation findings, and a comparison of prior approaches, suggest that the developed mode has a considerable increase in image processing attack robustness, making it ideal for copyright protection applications.