Embodiments of a system for food quality image classification are disclosed for classifying food items as either damaged or not damaged, and if damaged, the type of damage. Damage classification may be performed in multiple stages, for example, RGB data may be used in a first classification, and HSV data may be used in a second classification, such as whether damage is from bruising or decay. Blurring may improve determination accuracy, and an ongoing feedback and training loop may improve accuracy results over time.