PangaFlyIMG

The identification and counting of fish are important tools for managing the stock, production, and marketing of aquaculture fish. In commercial establishments, the counting of fingerlings has traditionally been done manually, which can cause stress to the animals and to the labor, in addition to low precision. Automation using computer vision and deep learning models is being increasingly explored. In machine learning for fish detection and counting, a dataset with images of specimens is needed to train the neural network. This dataset with species from the Panga family will be relevant in studies involving learning and training on different types of neural networks for counting and detection. The images were captured at a fish retailer. The company provided the location, the container, and the specimens, with an average length of 3.5 cm. 10 Panga fry were placed in a water container with a blue bottom, measuring 40 cm in diameter, 20 cm in height, and with a total capacity of 25 liters. The camera used to capture the images was fixed at the top, at a height of 60 cm from the container. The camera used was an iPhone XR with 12 megapixels and a resolution of 4608 × 2592 to capture the images. A total of 1,000 images were captured. The images present in the dataset are comprehensive, some have obstructions, low sharpness, and the water used in the container comes from the very recirculation system where the fry were raised, making the database more robust and closer to reality.

鱼类识别与计数是水产养殖鱼类种群存量管理、生产及营销的关键支撑手段。传统商业场景中，鱼苗计数多依赖人工操作，该方式不仅计数精度偏低，还会对水产动物与操作人员造成应激反应。目前，基于计算机视觉（Computer Vision）与深度学习（Deep Learning）模型的自动化计数方案正受到越来越广泛的研究与应用探索。在面向鱼类检测与计数的机器学习任务中，需通过包含鱼类样本的图像数据集来训练神经网络（Neural Network）。本数据集涵盖潘加科（Panga family）鱼类物种的样本图像，其将在面向各类神经网络的计数与检测学习及训练相关研究中具备重要应用价值。图像采集自某鱼类零售商户，该商户提供了拍摄场地、容器与实验样本，样本平均体长为3.5厘米。实验中将10尾潘加鱼苗放置于蓝色底面的水体容器中，该容器直径40厘米、高度20厘米，总容积25升。图像采集所用相机固定于容器正上方60厘米高度处，机型为iPhone XR，搭载1200万像素传感器，拍摄分辨率为4608×2592。本次采集共获取图像1000张。本数据集内的图像场景类型丰富，部分样本存在目标遮挡、清晰度偏低等问题；且容器内水体直接取自鱼苗原养殖的循环系统，这使得数据集更具鲁棒性，更贴近真实应用场景。