2026-10-14T20:03:07Z https://www.scidb.cn/oai

10.57760/sciencedb.25907 2026-10-08T09:27:57Z

2026-10-08 BDryFish: An Image Dataset for Dried Fish Classification in Bangladesh doi:10.57760/sciencedb.25907 en BDryFish is an image dataset of dried fish (shutki) as it is sold in Bangladesh. It contains 5,778 smartphone photographs of 22 market trade categories, taken on two field days, 14 and 21 May 2025, at two wholesale markets in Dhaka, Bangladesh, namely Karwan Bazar (23.7517 N, 90.3943 E) and Rayerbazar Shutki Arot in Mohammadpur (23.7456 N, 90.3587 E). Each category was photographed in two presentations. Single images (2,981) show isolated specimens laid on a plain sheet, and Bulk images (2,797) show the dense heaps in which the product is displayed for sale. Each category has between 210 and 345 images.The photographs were taken by hand, about 30 to 50 cm from the product and under natural light, with an Apple iPhone 15 (4,533 images), an Apple iPhone 12 (241 images) and a Samsung Galaxy F22 (1,003 images), each used with its main and ultra wide cameras; the flash fired on 700 images. At Karwan Bazar, the stalls of two vendors were photographed between 11:23 and 11:59 on 14 May (509 images of two categories). At Rayerbazar Shutki Arot, every stall, roughly 24 in all, was visited and each variety not yet recorded was photographed. Photographs were usually taken in short bursts, so consecutive frames often show the same specimen or heap from almost the same viewpoint.Of about 7,000 photographs, 5,875 were kept after screening by eye for blur, poor exposure, duplication and unrelated objects. A SHA-256 comparison then removed 96 byte identical copies in 87 groups, keeping one copy of each, and one empty file was removed, leaving 5,778 images. HEIC photographs from the iPhones were converted to JPEG with ImageMagick, and 2,531 frames were cropped by hand to remove distractions at their edges, so the full resolution images are the original JPEG files as kept after curation, not unprocessed sensor data. The resized copy was made from them with Python and Pillow by applying the EXIF orientation to the pixels, scaling the long side to 640 px with Lanczos resampling and saving at JPEG quality 95, without padding, cropping or augmentation. Files are named <Class>_<View>_NNNN.jpg and numbered in order of capture.The class labels are the Bangla trade names given by the sellers. English and scientific names are the closest matches found in online sources, checked against FishBase and the Marine Biodiversity Portal of Bangladesh, and are not taxonomic identifications. The categories combine species with processing state and cut. Hilsa and Slice Hilsa are one species, whole and sliced; Sun-Dried Swamp Barb and Fermented Swamp Barb are one species, sun dried and fermented; the two hairtail categories are sold as size grades; Chouka has no established scientific identity; and Small Shrimp is a crustacean.The record contains four files. BDryFish-v4-raw.zip (17.51 GB) holds the 5,778 images at full resolution, from 1.1 to 89.9 megapixels (long side 1,372 to 10,948 px). BDryFish-v4-resized640.zip (874.95 MB) holds the same images with a long side of 640 px; their dimensions vary because the aspect ratio is preserved. BDryFish-v5-metadata.zip (1.14 MB) holds the tables, split, checksums, scripts and documentation described below. README.md (8.92 KB) describes the record and replaces the README files inside the two image archives, which date from version 4. The three archives unpack into one folder, BDryFish-v4/, in which the images sit in raw/ and resized640/, each divided into 22 category folders and then into Single and Bulk folders. All files use common formats (JPEG images, UTF-8 CSV tables, JSON, Markdown and a Python 3 script) that open with standard software.The table metadata/images.csv has 5,778 rows, one per image, and 28 columns. They give the file name, category and presentation; the market; the capture date and time in Dhaka local time (UTC+6), with fractional seconds; the device, the lens (main or ultra wide, inferred from the recorded aperture) and the f-number; whether the flash fired; the EXIF orientation of the full resolution file; the width and height of both files in pixels; whether the full resolution frame matches a native output size of the phone, which marks cropped frames; capture session identifiers for gaps of 2, 5, 10 and 30 seconds; the session and threshold used by the recommended split; the split assignment (training, validation or test); SHA-256 checksums of both files; and the file name in version 1. The table metadata/classes.csv has 22 rows and 11 columns giving image counts, the Bangla, English and scientific names, processing state, cut form, any category of the same species and a note. The file splits/canonical_splits.csv (5,778 rows) holds the recommended split, metadata/removed_files.csv (97 rows) lists the files removed during curation, metadata/duplicates.json lists the 87 groups of identical copies, and metadata/filename_map.csv maps version 1 file names to the current ones. The folder checksums/ holds SHA-256 manifests for every file, and code/ holds bdryfish_tools.py, which rebuilds the resized images, the capture sessions and the split and verifies the checksums.One Bulk image (Indo-PacificKingMackerel_Bulk_0110.jpg) has no EXIF record, so its capture time, device, lens and flash state are empty and its market is inferred from its position in capture order. The iPhone photographic style (Standard, Vibrant or Rich Contrast) and the number of distinct specimens and heaps were not recorded, and cropping or conversion removed the lens model tag from 846 files. In raw/, 1,482 images are upright only when their EXIF orientation is applied, which image viewers do automatically but many programming libraries do not. Fifteen Galaxy F22 frames exceed the 48 megapixels of its main camera, as written by the phone's camera application. Because the labels rest on the sellers' naming and dried products are hard to identify, some labels may be wrong. Because frames were taken in bursts, a random split places near identical images in both training and test sets, so users should use the recommended split, which assigns whole capture sessions to one partition; at a 10 second threshold, 9.7% of Single and 26.5% of Bulk test images still share a session with training images, because some categories were photographed in a few long runs.Version 5 adds the metadata archive and a corrected README; the images are unchanged from version 4. The data are released under CC BY 4.0, and the scripts are also available at https://github.com/TanvirAnjumApurbo/BDryFish-tools. Dried Fish Classification; Bangladeshi Fish Dataset; Fish Image Recognition; Shutki; Market Trade Categories; Capture Sessions Tanvir Anjum Apurbo Kh Faysal Ahammed Hafizur Rahman Md. Abdullah PUBLIC https://creativecommons.org/licenses/by/4.0/ dataset Science Data Bank