Comparative Analysis of Long Short-Term Memory and Gated Recurrent Unit Models for Chicken Egg Fertility Classification Using Deep Learning ^†

Abstract

This study explores the application of advanced Recurrent Neural Network (RNN) architectures—specifically Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU)—for classifying chicken egg fertility based on embryonic development detected in egg images. Traditional methods, such as candling, are labor-intensive and often inaccurate, making them unsuitable for large-scale poultry operations. By leveraging the capabilities of LSTM and GRU models, this research aims to automate and enhance the accuracy of egg fertility classification, thereby contributing to agricultural automation. A dataset comprising 240 high-resolution egg images was employed, resized to 256 × 256 pixels for optimal processing efficiency. LSTM and GRU models were trained to discern fertile from infertile eggs by analyzing the sequential data represented by the pixel rows in these images. The LSTM model demonstrated superior performance, achieving a validation accuracy of 89.58%, significantly surpassing the GRU model (66.67%). Compared to classical methods such as Decision Tree (85%), Logistic Regression (88.3%), the LSTM model demonstrated superior performance, achieving a validation accuracy of 89.58%, significantly surpassing the GRU model (66.67%). Compared to Decision Tree (85%), Logistic Regression (88.3%), SVM (84.57%), K-means (82.9%), and R-CNN (70%), the LSTM model achieved the highest classification accuracy. Unlike classical machine learning approaches that rely on handcrafted features and predefined decision rules, LSTM effectively learns complex sequential dependencies within images, improving fertility classification accuracy in real-world poultry farming applications. In contrast, GRU models, while more computationally efficient, may struggle with generalization under constrained data conditions. This study underscores the potential of advanced RNNs in enhancing the efficiency and accuracy of automated farming systems, paving the way for future research to further optimize these models for real-world agricultural applications.

1. Introduction

Classifying chicken egg fertility is crucial for the poultry industry [1] as it directly impacts production efficiency and profitability [2]. Traditional methods like candling, which illuminates eggs to visualize embryos, require skilled personnel and are only sometimes accurate or feasible for large-scale operations due to their manual nature [3,4]. Recent advancements in machine learning [5,6], particularly image classification with Recurrent Neural Networks (RNNs) [7], offer promising solutions for automating and enhancing this process. Classical machine learning approaches [8] such as Decision Tree and Logistic Regression rely on manually extracted features, limiting their adaptability to variations in image quality and lighting conditions. While models like SVM and K-means clustering improve upon purely statistical classifiers, they still struggle with capturing complex spatial relationships. LSTM and GRU, by contrast, learn long-range dependencies within images, making them more practical for real-time, automated fertility classification in poultry farming. RNNs, especially Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRUs) are adept at processing the sequential and spatial data [9] critical to identifying embryonic development in egg images [10,11].

This study employs LSTM and GRU to address the shortcomings of traditional and manual egg fertility assessments [12,13]. It leverages their capability to handle complex data dependencies without the issues of simpler RNNs, such as vanishing gradients. By analyzing each row of pixels as a sequence, these models effectively aim to discern fertile from infertile eggs.

The innovation of this research lies in applying LSTM and GRU to the field of embryology within poultry science—a novel approach given the limited exploration of these advanced neural networks in such a context. The objectives are to develop, fine-tune, and compare these models on a dataset of 240 resized egg images to determine which model best balances accuracy with computational efficiency. Ultimately, this study seeks to enhance poultry farming technology and contribute to agricultural automation by showcasing the practical application of sophisticated neural networks.

2. Background and Related Work

2.1. Technological Background

Recurrent Neural Networks (RNNs) are a class of neural networks designed to handle sequential data [9] by introducing loops that allow information to persist within the networks [14]. This characteristic makes RNNs ideal for involving time series or sequential data, such as language modeling, speech recognition, and sequential image analysis, which is the basis of this study on egg fertility classification [15]. A specific type of RNN, Long Short-Term Memory (LSTM), was developed to mitigate the vanishing gradient problem that can limit the ability of traditional RNNs to learn long-range dependencies within sequences [16]. An LSTM unit has three gates: an input gate, a forget gate, and an output gate [17]. These gates collectively decide what information to discard and what to retain, which helps maintain the flow of relevant information through the network [18].

Gated Recurrent Units (GRUs) simplify the LSTM architecture by combining the input and forget gates into a single “update gate” and merging the cell and hidden states [19]. This simplified structure reduces computational complexity while achieving performance comparable to LSTMs in many applications. In situations where computational efficiency is paramount, GRUs offer a compelling alternative to LSTM, particularly in cases with limited data availability or when real-time performance is needed [20].

This study extends the application of LSTM and GRU to embryology within poultry science, a relatively unexplored domain for advanced RNN architectures. Unlike conventional egg fertility assessment methods, such as candling, which are labor-intensive and prone to inaccuracies, this research investigates the potential of LSTM and GRU for automating the classification of chicken egg fertility based on sequential image data analysis.

2.2. Review of Related Work

Previous studies have explored various image processing techniques for egg fertility assessment [2], with methods ranging from traditional image thresholding to more complex machine learning algorithms [3,21,22]. However, these studies primarily focus on external characteristics and rarely analyze internal structural patterns critical for embryonic development detection. Techniques like support vector machines (SVMs) and Decision Trees have been employed [3,23]. However, they often require hand-crafted features and may not capture the deeper patterns within the data that deep learning models can efficiently identify.

Deep learning, particularly Convolutional Neural Networks (CNNs), has shown promise in many domains of agricultural image processing. However, RNNs in agriculture remain limited, specifically for tasks like egg fertility detection. Some studies have explored RNNs in crop yield prediction and plant disease detection using sensor data or temporal satellite images [24]. Studies utilizing RNNs to monitor animal behavior patterns [25] or predict livestock outcomes [26] further validate the applicability of RNNs in varied contexts within agriculture [27]. These applications demonstrate the suitability of RNNs for agriculture, where sequential patterns in data play a critical role. However, few applications address sequential image data for poultry fertility classification [28].

This gap in the literature signifies an opportunity for innovative research in applying LSTM and GRU models for egg fertility classification, a crucial task in the poultry industry. This research aims to address this gap by comparing the effectiveness of these advanced RNN architectures in classifying chicken egg fertility based on embryonic development detected in egg images.

3. Method

3.1. Dataset and Image Preprocessing

The research involves a dataset of 240 high-resolution images of chicken eggs, precisely divided into two classifications: fertile and infertile [3] (sample as shown in Figure 1). The images were captured using the candling process [3], with a 5000-lumen LED flashlight and a 5MP smartphone camera in a dark room to control lighting. The camera was positioned 20 cm above the egg, with automatic settings (focus, exposure, resolution) to ensure consistency. Further details on the camera setup could clarify potential biases in data collection. This categorization was performed through expert assessments, ensuring accuracy and reliability for model training. This balanced dataset prevents biases in model training and performance evaluation.

All images in the dataset were resized from their original dimensions of 665 × 545 pixels [1] to a uniform size of 256 × 256 pixels. This resizing is essential for standardizing input data and enhancing the neural networks’ processing efficiency. It also helps reduce computational demand, thus speeding up the training process. The selected dimension (256 × 256) strikes a balance between maintaining sufficient image detail for detecting fertility features and optimizing computational resources.

3.2. Model Architectures

LSTM and GRU models (as shown in Figure 2) are designed to process the spatial information in the images sequentially [29]. Each image is reshaped such that each row of 256 pixels becomes a timestep in the sequence, with 256 rows each containing 765 features derived from RGB values. Each row consists of 256 features derived from the image’s RGB values, providing the network with enough spatial information to detect relevant patterns and dependencies for egg fertility classification.

Figure 2 details the configuration of each layer in both the LSTM and GRU models. The models utilize feature extraction via CNN (Convolutional Neural Networks) to pre-process the input image, followed by LSTM and GRU layers for sequential processing. The key components include:

LSTM model [30]: The LSTM layer, with 128 units [31,32], captures long-range temporal dependencies in the sequential data, helping to recognize patterns in embryonic development. The model concludes with a single output layer of binary classification (fertile vs. infertile eggs).

GRU model: The GRU layer, also with 128 units [33,34], simplifies the architecture by combining the forget and input gates into a single update gate, maintaining efficiency while still learning complex temporal dependencies. Like the LSTM, it uses an output layer for binary classification.

Both models aim to capture temporal dependencies in sequential data, with the LSTM excelling in handling long-range dependencies and the GRU offering a more computationally efficient alternative.

3.3. Training Procedures and Evaluation Metrics

Training for models is executed using the Adam optimizer, known for its adaptive learning rate capabilities, which help manage sparse gradients effectively [35,36]. The Adam optimizer adjusts the learning rates based on the first- and second-moment estimates of the gradients. The update rules for each parameter θ in the model are given by Equations (1)–(4).

$$m_t={\beta }_1{m}_{t-1}+\left(1-{\beta }_1\right)g_t$$

(1)

$${\widehat{m}}_t={\displaystyle \frac{m_t}{1-{\beta }_1^t}}, {\widehat{v}}_t={\displaystyle \frac{v_t}{1-{\beta }_2^t}}$$

(2)

$$v_t={\beta }_2{v}_{t-1}+\left(1-{\beta }_2\right)g_t^2$$

(3)

$${\theta }_t={\theta }_{t-1}-{\displaystyle \frac{\eta }{\sqrt{{\widehat{v}}_t+ϵ}}}{\widehat{m}}_t$$

(4)

where $g_t$ is the gradient of the loss function with respect to θ at time step t, m_t and v_t are the first- and second-moment estimates, and ${\widehat{m}}_t$ and ${\widehat{v}}_t$ are the bias-corrected versions of the first- and second-moment estimates. The learning rate η and a small constant ϵ are used to prevent division by zero. This update rule ensures efficient parameter adjustments during training, helping to optimize model performance.

The binary cross-entropy loss function is used for training, which is particularly suited for binary classification tasks [37]. It measures the difference between the true labels y and the predicted probabilities $\widehat{y}$ produced by the model. The binary cross-entropy loss is computed as Equation (5).

$$\mathcal{L}=-{\displaystyle \frac{1}{N}}\sum _{i=1}^N\left[y_i\log \left({\widehat{y}}_i\right)+\left(1-y_i\right)\log \left(1-{\widehat{y}}_i\right)\right]$$

(5)

where $\mathcal{L}$ represents the loss, N is the batch size, ${\widehat{y}}_i$ is the true label for sample i, and ${\widehat{y}}_i$ is the predicted probability for the same sample. The binary cross-entropy loss penalizes incorrect predictions more heavily, guiding the model to output probabilities closer to the true labels as training progresses.

The models are trained for 100 epochs with a batch size of 32, allowing the models ample opportunity to learn from the dataset thoroughly. This process ensures the neural networks capture nuances needed for egg fertility classification. The training process includes a 20% validation split, which periodically evaluates the model against unseen data, thus preventing overfitting and ensuring the models are generalizable. Model performance is assessed using accuracy, which measures (Equation (6)) classification effectiveness, and loss (binary cross-entropy), which indicates the models’ predictive confidence and how well the probability outputs align with correct labels during training and validation.

$$Accuracy={\displaystyle \frac{\sum _{i=1}^{N}1\left({\widehat{y}}_i=y_i\right)}{N}}$$

(6)

where $1\left({\widehat{y}}_i=y_i\right)$ is an indicator function that equals 1 when the predicted label yˆ_i matches the true label y_i, and 0 otherwise. This provides a clear indication of how well the model is performing in terms of classification.

4. Results and Discussion

4.1. Performance of Each Model

The LSTM model showcased promising results with a peak training accuracy of approximately 95%. However, the validation accuracy plateaued at around 70%, suggesting the model might be overfitting the training data Figure 3a,b. The training loss showed a significant decline, which stabilized in later epochs, but the validation loss did not parallel this trend, hinting at possible generalization issues.

Similarly, the GRU model demonstrated high training accuracy, nearing 98%, but its validation accuracy hovered between 65% and 67% (Figure 3c,d). Like the LSTM, the GRU model’s training loss decreased initially but leveled off, while the validation loss remained higher and more volatile. This pattern further supports the occurrence of overfitting, with the model learning the training data well but failing to generalize effectively to unseen data.

4.2. Comparative Results of LSTM and GRU

This section evaluates and compares the performance of the LSTM and GRU models based on their validation results, focusing on accuracy, loss, and a detailed examination of their confusion matrices.

The LSTM model achieved an impressive validation accuracy of 89.58% with a loss of 1.1691. This high accuracy indicates the model’s strong capability to generalize from the training data to unseen data, suggesting effective learning and adaptation to the features crucial for classifying egg fertility. The performance surpasses traditional machine learning methods, with SVM [23] achieving 84.57%, K-means [38] reaching 82.9%, and R-CNN [21] models performing at 70% accuracy. These results underscore the effectiveness of the LSTM in leveraging sequential data for this task.

In contrast, the GRU model recorded a lower accuracy of 66.67% and a significantly higher loss of 12.6634. The increased loss and reduced accuracy highlight challenges in the model’s convergence or potential overfitting to less representative features of the dataset, which could impair its generalization capabilities. While the GRU’s simpler architecture provides computational efficiency, it does not capture the necessary temporal dynamics as effectively as the LSTM, resulting in poorer performance than LSTM and other traditional methods.

The performance nuances of each model can be further revealed through their confusion matrices in

Table 1. Confusion matrix of LSTM and GRU: shows the number of true and false predictions for fertile and infertile eggs by each model.

Confusion Matrix	LSTM Prediction		GRU Prediction
	Positive	Negative	Positive	Negative
Actual Positive	16	13	19	7
Actual Negative	8	11	17	5

. The LSTM model shows a balanced ability to identify both fertile and infertile eggs, with some room for reducing false positives and false negatives. This indicates that while the model performs well, slight improvements in tuning could further optimize its performance in distinguishing between the two classes. The GRU model shows more true positives (19) for fertile eggs and more false positives (17). This suggests that the GRU model is more aggressive in predicting fertility, leading to a higher misclassification rate, where infertile eggs are incorrectly classified as fertile.

The superior performance of the LSTM model in both accuracy and loss is attributed to its complex internal gating mechanisms, which efficiently manage long-range dependencies and mitigate significant overfitting better than the GRU model. The simpler architectural design of the GRU, while beneficial in terms of computational efficiency, might not have captured the necessary temporal dynamics as effectively as the LSTM, resulting in its poorer performance. This analysis suggests that while the GRU model may excel in scenarios requiring rapid processing with acceptable accuracy, the LSTM is preferable for tasks where precision and reliability are paramount.

Furthermore, when compared with other methods, such as Decision Tree (85%), Logistic Regression (88.3%), SVM [23] (84.57%), K-means [38] (82.9%), and R-CNN [21] (70%), the LSTM model outperforms these traditional machine learning approaches by a significant margin, emphasizing its superiority for this specific application.

5. Conclusions

This study compared LSTM and GRU models for classifying chicken egg fertility based on embryonic development in egg images. The LSTM outperforms the GRU in accuracy and generalization, effectively capturing complex temporal dependencies. While the GRU is more computationally efficient, it struggles with overfitting and generalization due to the limited dataset. Achieving 89.58% accuracy, the LSTM model shows strong potential for automating egg fertility classification in poultry farming. Unlike traditional methods such as Decision Tree (85%), Logistic Regression (88.3%), SVM (84.57%), K-means (82.9%), and R-CNN (70%), which rely on manual feature extraction and predefined statistical assumptions, LSTM can automatically learn sequential dependencies within images. This makes it more adaptable to different imaging conditions and more suitable for precision agricultural applications where accuracy and scalability are essential. Future work should focus on improving LSTM’s generalization through data augmentation, diverse datasets, and hybrid models. Expanding these models to other agricultural domains could further enhance deep learning’s role in agricultural efficiency.