In this paper, we propose the implementation of optimization based Extreme Learning Machine using the Finite Newton-Armijo method (FNELM) for digital watermarking of color images. The Finite Newton-Armijo method for ELM convergence is established and mathematically justified. The ELM is a regularized algorithm based on generalized single-hidden-layer feedforward neural networks (SLFNs). However, a hidden layer in the context of ELM, this is commonly referred to as "feature mapping," so you do not have to tune it again and again. This paper includes the embedding and extraction processes of a watermark, a logo, or a sequence of numbers in a color image with the help of the said machine. As a result, the ELM provides better generalization performance, the approach tends to deliver superior performance by simultaneously reaching the lowest training error and the smallest norm of the output weights. The efficacy of the FNELM algorithm is measured using the Peak Signal-to-Noise Ratio (PSNR), Bit Error Rate (BER), and other qualitative parameters. Retrieval rate R and normalized correlation coefficient (NC) on signed images and, further, on attacked images created by applying malicious attacks over them. The results obtained are compared with some existing methods, and the proposed method outperformed them in comparison.