Multistage Fish Eyes Dataset

The Multistage Fish Eyes Dataset (MFED) is a comprehensive image dataset constructed to facilitate research in fine-grained visual classification for fish freshness assessment under real-world conditions. The dataset comprises a total of 4,800 images of fish eyes. The specimens used for this dataset include two common freshwater species: Cyprinus carpio var. Fanli (Rice flower fish) and Carassius auratus (Crucian carp).
To accurately capture the biological degradation process, the freshwater fish specimens were preserved at a consistent storage temperature of 4℃ throughout the study. The continuous data acquisition process lasted for six days. To enhance the generalization ability of models trained on this data and to intentionally introduce real-world environmental interference, images were systematically captured under four varying illumination and background setups:
·White light with a cutting board background.
·White light with an A4 paper background.
·Warm yellow light with a cutting board background.
·Warm yellow light with an A4 paper background.
For each specimen, images of both the left and right eyes were taken daily from five different shooting angles. The camera was positioned at a distance of 20 - 30 cm using a 3.7x zoom. This methodology yielded a total of 40 images per fish per day across the different environments.
The dataset is perfectly balanced, featuring exactly 1,600 images in each of the three freshness categories. These categories are strictly defined by storage time and distinct visual biological indicators:
·Highly Fresh: Specimens stored for ≤ 2 days, visually characterized by almost no corneal opacity and a cleared pupil.
·Fresh: Specimens stored for > 2 days and ≤ 4 days, visually characterized by slight corneal opacity and blurred pupil margins.
·Not Fresh: Specimens stored for > 4 days, visually characterized by severe corneal opacity and pupil deformity.