The success of classical machine vision algorithms varies greatly depending on multiple factors and often requires the input of a computer vision expert and an implementation in an environment where lighting is perfectly controlled to avoid unwanted reflections and specular highlights that could degrade performance of tasks such as object recognition. However, as lighting conditions can constantly change in human-robot collaborative settings, these implementations that require perfect conditions are not always applicable. The main goal of this project is to develop a method to automatically enhance image quality for object recognition algorithms in industrial settings with little or no control on lighting conditions. The developed method will rely on machine learning to train a neural network. To train the network, an annotated database of common industrial objects in various lighting conditions that are manipulated by light to medium duty robot arms will have to be produced, as commonly available annotated databases for object classification do not contain this type of data. Then, the network will be trained to remove lighting artifacts such as specular highlights and produce images suitable for object recognition and localization. The performance of the method will be measured in grasp point localization scenarios.