BDryFish: An Image Dataset for Dried Fish Classification in Bangladesh

The 22 classes are market trade categories, not taxonomic identifications. They combine species with processing state (Fermented Swamp Barb vs Sun-Dried Swamp Barb, both Puntius chola) and cut form (Hilsa vs Slice Hilsa, both Tenualosa ilisha), and they include two congeneric hairtail categories distinguished by size grade (Giant Largehead Hairtail, LargeHead Hairtail, both Trichiurus lepturus). Three pairs of classes are therefore the same species presented or processed differently.

This is a deliberate property: it reflects how the commodity is actually traded and priced, and it makes the dataset usable for fine-grained tasks that species-level labels cannot express.Contents. Two archives are provided. They contain the same 5,778 images under identical filenames and folder structure — resolution is the only difference between them.BDryFish-v4-raw.zip (17.5 GB) — the original camera files at full resolution (1.1–89.9 MP; long side 1,372–10,948 px, median 12.2 MP), EXIF intact, no GPS coordinates.BDryFish-v4-resized640.zip (0.85 GB) — every image resized so its long side is 640 px.

Aspect ratio preserved; no padding, no cropping, no augmentation. Lanczos resampling, JPEG quality 95, 4:4:4 chroma, EXIF orientation applied to the pixels so all images are stored upright.       Both unpack to BDryFish-v4/<raw|resized640>/<Class>/<Bulk|Single>/<Class>_<View>_NNNN.jpg. The class folder name is the label.

Per-class counts range from 210 to 345 images (2,797 Bulk, 2,981 Single). A README in each archive.Version 4. This version supersedes all earlier versions of byte file and 96 redundant byte-identical duplicate copies were removed, and an image-orientation defect in the earlier resized set — in which EXIF rotation was not applied before resizing, leaving roughly a quarter of the derivative images rotated relative to their originals — has been corrected.

The earlier fixed-size derivative set, which was letterboxed to 640×640 with black bars, has been replaced with size. Image counts differ from earlier versions, so please state the version you used.Intended uses. Fine-grained image classification; robustness and domain-shift evaluation across presentation conditions; texture and clutter recognition; food-product identification, market surveillances for a commodity sector that is economically significant in South Asia and poorly represented in existing image resources.Limitations.

Collection is confined to two markets in a single city over two capture days, so seasonal, regional and vendor-specific variation is not represented. Labels are market trade categories verified by tory taxonomic identification. Classes are moderately imbalanced.

Photographs contain product only; no identifiable persons appear in the images and no location coordinates are recorded in the files.