According to World Corrosion Organization (WCO), the estimated annual cost of damage due to corrosion across the globe is approximately US$2.5 trillion which contributes 3–4% GDP of developed countries. Minimizing losses due to corrosion and to ensure longer life of structures is thus a major concern. Paper presents a technique employing bag of features (BoF) for underwater structural corrosion recognition. BoF methods are based on an unorganized grouping of image features and it is conceptually simpler than various other alternatives. The model is trained on three labeled datasets corroded, un-corroded, and damaged obtained from underwater structures along the Gomti River in Lucknow. Dataset of around 2000 images is used to train the model. Trained prototype BoF learning model is capable of efficiently classifying pure and corroded images and achieves an accuracy of 82.38% demonstrating the feasibility of this method. The technique proposed and its deployment on handheld and autonomous devices provide an efficient and intelligent method for underwater structure corrosion recognition.