The current study indicates that deep learning may be effectively used in applications including picture categorization, image identification, and object recognition by using several CNN architectures. On altered and/or bigger datasets, cost-effective picture classification is accomplished, and enhanced image feature mapping is derived from related images in text metadata using CNNs. Given the limited association between feature labels and comparable (and/or unrelated) pictures, employing feature map representations is demonstrated to be cheaper and quicker, but it does not increase the quality of the image classifications, suggesting that this technique is not ideal for assessing quality. However, using the newly acquired learnt weights, the findings of the current study may inspire further research into alternative counterfeit detection methods. Overall, our study shows that metadata sampling and categorization need a highly disciplined scaling model, which can be scored by using a pre-trained model, and which may be further developed in future phases.