A Hybrid Image Training Dataset of 12 Freshwater Ornamental Fish Species

This dataset was created to train a fish recognition model for 12 common freshwater aquarium species. The data was gathered using three distinct methods. First, existing public datasets were downloaded and combined from online sources like Kaggle. Second, a Python web-scraping script and manual downloads from Google Images were used to collect additional pictures. The third method was primary self-collection, conducted specifically for the Cherry Barb class. To gather this data, a Sony ZV-1F camera and a smartphone were used to photograph live Cherry Barbs in a home planted aquarium, which introduces real-world visual challenges like water refraction and plant occlusions. The dataset is organized into three folders to enable easy reuse: "Original Cropped Images" contains centered specimens ready for classification models, "Raw Aquarium Shots" holds the full, unedited tank photos of the Cherry Barbs for object detection tasks, and "Augmented Images" provides an artificially expanded dataset for deep learning. Researchers can use this hybrid data to train computer vision models and test how well they perform on clear web-sourced images compared to real-world aquarium photos.

本数据集专为12种常见淡水观赏水族物种的鱼类识别模型训练而构建。数据采集采用三种不同方式：其一，从Kaggle等在线平台下载现有公开数据集并进行整合；其二，借助Python网络爬虫脚本与谷歌图片手动下载的方式，补充收集图像素材；其三，针对樱桃鲫（Cherry Barb）类别开展专属的原始自主采集工作。本次采集使用索尼ZV-1F相机与智能手机，在家庭水草缸中拍摄活体樱桃鲫，该场景会引入真实世界的视觉挑战，如水折射与水草遮挡等。
该数据集划分为三个文件夹以提升复用便捷性：“原始裁剪图像（Original Cropped Images）”存放已完成居中裁剪的样本图像，可供分类模型直接使用；“水族实拍原图（Raw Aquarium Shots）”收录樱桃鲫的完整未编辑缸内照片，适用于目标检测任务；“增强图像（Augmented Images）”提供经人工扩展的数据集，用于深度学习训练。研究人员可借助该混合数据集训练计算机视觉模型，并对比模型在网络开源清晰图像与真实水族场景实拍图像上的性能表现。