Image Quality Aware Deep Learning Networks for Budgerigar Gender Recognition
Michael A. Hsiao, Kyle Mooney, Jinhui Wang · 2025
Although traditional human image recognition suggests that higher resolution images will yield better results than lower resolution images because they hold more details, the results of this study prove that low-resolution images may have their own niche in the training of Artificial Intelligence (AI) models for more accurate and reliable image recognition. The purpose of this study is to investigate the impact of image resolution on the performance of an Artificial Neural Network (ANN), by case study on budgie gender identification. This study will be an important reference for the training of color-oriented models by encouraging low-power and small-size data storage due to low-resolution images, decreasing cost and resources of AI applications. In the study, multiple data sets of images of budgie ceres with varying image resolutions, ranging from 22×22 to 256×256 pixels, are used to train and validate neural networks and test the reliability. It finds that the peak and end accuracy of the models are both right skewed with respect to the pixel size distribution and high accuracy. High reliability can be associated with a low image resolution, 22×22, 25×25, and 32×32 pixels image data set. The highest accuracy is obtained with an optimized ratio of a number of pixels to a number of images in a training dataset to create a stable learning rate and prevent overfitting or underfitting too rapidly.