Prediction of live weight of hens using non-stationary thermal imaging and deep learning algorithms

1. This study investigated predicting the live weights of 48-week-old Lohmann Sandy breed laying hens using non-stationary thermal images and CNN image processing and deep learning algorithms. A training set was created using a specially designed platform, consisting of 2,116 thermal images, with at least 10 frames for each of the 203 laying hens. The test set consisted of 50 images from four hens, which were completely excluded from the training phase. This allowed the deep models to reflect their true ability to predict novel live weights for laying hens, rather than memorising patterns from known subjects.2. To ensure data consistency and quality, thermal images were resized to 224 × 224, 299 × 299 and 380 × 380 pixels before being added to the deep learning framework. Because resolution varies depending on camera settings and distance, standardisation of image dimensions was used to ensure consistency in the neural network's input layer.3. Standardised thermal images of laying hens were built on a convolutional neural network (CNN) architecture and used the EfficientNetB1, Xception, DenseNet121 and InceptionResNetV2 deep learning algorithms to predict live weight. The training process used Google's comprehensive deep learning framework, TensorFlow and its high-level API, Keras, which simplifies model building and training workflows.4. The test set prediction performance for live weight of laying hens using DenseNet121-Linear were RMSE (0.046), MAPE (2.154%), ME (0.022%), RAE (2.739%) and R² ;(0.71), when using non-normalised data.5. The study demonstrated that the DenseNet121 deep learning model can effectively predict laying hen weights using only thermal imaging and it is anticipated that this will lead to numerous practical applications in the poultry industry.