Abstract

This study aims to address the limitations of traditional methods for measuring duck body dimensions, which are essential for assessing growth, health, and meat quality in poultry farming. Traditional techniques, relying on manual measurements, are time-consuming, labor-intensive, and prone to human error, often causing stress to the animals. With the advancements in computer technology and artificial intelligence, particularly in the field of computer vision, there is potential to revolutionize how these measurements are conducted. Previous research has explored the application of machine learning and neural networks in poultry farming, achieving varying degrees of success. However, studies specifically focusing on ducks remain scarce. This research proposes a novel model combining Transformer and Convolutional Neural Networks (CNN) for predicting duck body dimensions. The study’s primary objectives include collecting a unique dataset that combines visual information with body dimension data of poultry, developing and optimizing the Transformer-CNN-based model, and evaluating its effectiveness and performance in predicting duck body measurements. The findings from this study are expected to contribute to enhancing breeding efficiency and economic benefits in poultry farming through improved measurement techniques.

Dataset Collection

We conducted a comprehensive data collection process at the breeding base of Linwu Shunhua Duck Industry Co., Ltd., located in Linwu County, China. Utilizing Intel RealSense depth cameras and RGB cameras, we captured the first dataset in this field that includes poultry body dimension information along with point cloud data and RGB images specifically for ducks. The dataset collection process was significantly optimized by the development of an integrated software solution designed to handle data acquisition and preliminary processing from multiple vision sensors simultaneously. This software dramatically increased the efficiency of the data collection process, ensuring that high-quality, precise data were gathered to support the development and validation of advanced predictive models for duck body dimensions. This dataset is expected to serve as a valuable resource for future research in the domain of poultry farming and computer vision applications.

• Related Software

Features Extraction

In the feature extraction phase of this study, we focused on processing the collected point cloud images to identify and extract critical features relevant to duck body dimensions. We employed specialized annotation software to perform 3D point labeling on the point cloud data, marking five key anatomical landmarks on each duck. These key points were selected based on their relevance to accurately describing the duck’s body dimensions.

To automate and enhance the precision of key point identification, we utilized the PointNet++ model, a well-established deep learning architecture designed for point cloud data. By training the PointNet++ model on our annotated dataset, we were able to reliably predict the positions of these key points in the point clouds.

Additionally, to further improve the model’s accuracy in predicting duck body dimensions, we calculated and incorporated the distances and angles between the key points as features. These geometric relationships provide valuable information about the overall form and structure of the duck, contributing to a more robust and accurate model. This feature extraction approach not only streamlines the data processing workflow but also enhances the predictive power of the final model, leading to more precise measurements and better support for applications in poultry farming.

Model & Experiment

Sorry! This part will be released after the publication of our paper.