A lightweight visual system for determining the gender of poultry based on artificial intelligence

In the bustling world of modern poultry farming, a simple biological fact drives complex economic decisions: male and female chicks require different diets to thrive. Males grow best on high-energy food to build weight, while females need high-protein meals to maximize egg production. For decades, farmers have relied on two main ways to tell the sexes apart after the chicks hatch. One method involves skilled workers manually inspecting the tiny cloaca, a task that is fast but prone to human error and fatigue. The other uses specialized machines that are accurate but prohibitively expensive and slow for large-scale operations. Both approaches struggle to keep up with the demands of modern, efficient farms, leaving a gap for a solution that is both affordable and precise.

This challenge has led researchers to explore artificial intelligence, specifically using computer vision to "see" the differences between male and female chicks. While some previous attempts have tried to analyze the sound of chirping or the texture of feathers, these methods often fail in the noisy, unpredictable environment of a real hatchery. Sound is easily drowned out by machinery, and feathers can obscure the subtle physical differences needed for identification. To solve this, a team of researchers has developed a new visual system that focuses directly on the chick's gonads, the internal organs that determine sex. By combining advanced image recognition with a specialized camera setup, they have created a tool that can identify the gender of a chick with high accuracy, offering a potential path toward more humane and efficient farming.

The researchers began by tackling the most fundamental hurdle: data. To teach a computer how to distinguish between a male and a female chick, they needed a vast library of clear images showing the specific physical traits of each. They collected over 1,200 high-resolution photographs of one-day-old yellow broiler chicks. In a careful, non-invasive process, specialists gently exposed the cloacal region of each chick to capture the gonads. The images revealed a distinct difference: male chicks display a small, circular protrusion, while female chicks appear relatively flat. However, capturing these images is difficult because the organs are hidden beneath soft down feathers, and the process often involves human fingers holding the chick, both of which can block the view or create confusing shadows. To overcome this, the team built the first standardized dataset of its kind in China, meticulously labeling each image as either male or female, creating a foundation for their artificial intelligence system.

With the data in hand, the team constructed a two-part system designed to mimic the way a human expert might approach the task, but with the speed and consistency of a machine. The first part acts as a locator. Using a well-known object detection technology, the system scans an image of a chick and instantly finds the exact spot where the gonads are located, ignoring the rest of the bird, the feathers, and the human fingers holding it. It then crops out just that small area, resizing it to a standard format. This step is crucial because it removes the visual clutter that often confuses other systems, ensuring the computer only looks at the relevant biological features.

The second part of the system is the classifier, the "brain" that makes the final decision. This component uses a specialized neural network, a type of computer program modeled after the human brain, to analyze the cropped image. The researchers designed this network to pay attention to two specific things: the overall shape of the area and the specific details within it. They integrated a mechanism that allows the system to focus on the most important parts of the image while ignoring the noise, much like how a human might squint to focus on a specific detail while blocking out the background. By combining deep layers of feature extraction with this dual focus, the system learns to recognize the subtle circular bump of a male or the flat surface of a female, even when the image is slightly imperfect.

When tested, the system performed with remarkable precision. In controlled experiments, the new system correctly identified the gender of a chick in more than 91 percent of cases. This was a significant improvement over using a standard detection model alone, which struggled with an accuracy of only 55 percent, highlighting how essential the specialized design was for handling the messy reality of real-world images. The researchers also compared their system against several other popular artificial intelligence models, including those based on complex transformer architectures and classic convolutional networks. Their system not only achieved the highest accuracy but also did so with fewer computational resources, making it faster and more suitable for the limited hardware often found in smaller farms.

To ensure the system was not just guessing but actually learning the correct biological features, the researchers used a technique to visualize where the computer was "looking" when it made a decision. The results showed that their system focused sharply on the gonadal region, specifically the protrusion or the flat area, whereas other models often got distracted by the surrounding feathers or the human fingers. This confirmed that the system was relying on the same anatomical cues that human experts use, rather than finding accidental patterns in the background. In a real-world trial set up in a commercial hatchery, the system ran continuously for six hours on modest, affordable hardware, processing about 11 to 12 chicks every second with consistent accuracy, regardless of changes in lighting or who was handling the birds.

The study concludes that this lightweight visual system offers a viable, low-cost alternative for the poultry industry. By automating the gender identification process with high accuracy, it promises to reduce the reliance on manual labor and expensive machinery, potentially lowering costs and improving animal welfare. The researchers have made their dataset and the source code for their system publicly available, inviting others to build upon this work. While the current system was tested on a single breed of chick under specific conditions, the success of this approach suggests a clear path forward for developing robust, automated tools that can adapt to the diverse and demanding environments of modern agriculture.