GAN-based motion blur elimination as a preprocessing step for enhanced AI-driven computer-aided camera monitoring in poultry and free-range farming in low-resource settings

Accurately determining poultry gender ratios is essential for assessing the economic value of free-range farming, particularly in resource-limited settings. Traditional and manual methods for gender identification are often labor-intensive, time-consuming, and prone to errors, especially in the early stages of poultry development. To address these challenges, this study introduces an automated approach that uses advanced machine learning techniques. Specifically, we propose a classification framework that integrates Convolutional Neural Networks (CNNs) with Generative Adversarial Networks (GANs) to enhance the accuracy of poultry gender identification. Our framework incorporates a novel GAN-based motion blur elimination method, which is broadly applicable to detect and classify moving subjects, including poultry. The proposed approach demonstrates a 98% accuracy in distinguishing between male and female birds at early growth stages by analyzing key features such as crown pixel measurements, feather gap analysis, and leg measurements. Furthermore, we conduct a comprehensive comparison of four segmentation models—UNet, ResUNet, ResUNet+, and a novel GAN-enhanced UNet—under varying motion blur conditions (80%, 50%, 30%, and 10%). Our results highlight the superiority of ResUNet+ over conventional models, achieving a peak Dice Coefficient of 91.2%, an Intersection over Union (IoU) of 86.7%, and the highest segmentation accuracy at reduced blur levels. These findings underscore the efficacy of deep learning-based approaches in advancing poultry gender classification while improving image quality in dynamic environments. • We introduce a classification framework for predicting the GAN-based motion blur elimination for the moving population as a preprocessing step. • We collected 1000 images, focusing on birds aged between 3 to 4 months. • We employed DeepLabV3 with Xception Backbone for semantic segmentation block to remove the background, which resulted in a pixel accuracy of 96.8%. • The CNN pixel classifier achieved a classification accuracy of 98%, showcasing the three important features of bird gender based on the crown pixel measurement, pixels between feathers, and measurement of legs at the early stage. • We present innovative methods for the gender detection of poultry birds, offering a faster and more efficient alternative to traditional detection approaches.