Dataset

Data Collection

We finalized the steak grades classified in this project to be USDA Select, Choice, Prime, and Wagyu steak. Wagyu steak is from a Japanese beef cattle breed. It is highly marbled and has a unique taste. Since Wagyu is a popular choice of steak but is not graded by USDA, we decided to set it as a separate grade. Regardless of their type (ribeye, strip, T-bone, etc.), steak pictures were collected from the internet to use as training data.

To improve the robustness of the model, we collected pictures from multiple angles with different backgrounds and lighting conditions. Besides standard top or side views of steak with a dark background, we collected images with backgrounds such as a piece of paper, a table with decorations, and a package with labels as seen in grocery stores. The pictures are collected from a variety of sources, therefore including different lighting conditions. Apart from the images with a single piece of steak, there are pictures of different numbers of steak with the same grade. More than 100 pictures were collected for each grade.

Grade - Select

Grade - Choice

Grade - Prime

Wagyu

Data Preparation

The collected dataset varies from formats to resolutions. All images are converted to png images for convenience, and the shorted side is resized to 512 pixels. The data augmentation techniques were applied to expand the population of the dataset, including rescale, hflip, vflip, and color jitter.

Rescale

Vertical flip

Horizontal flip

Color jitter

Examples of augmented dataset