A approach for intelligently classifying the state of chicken eggs based on deep associative neural networks is proposed. This method aims to automate the recognition and interpretation of chicken egg ovoscopy visualization results during incubation. The model of the associative autoencoder offers several advantages over traditional methods. For instance, the input image is pre-sized, and the pairs count "convolutional pooling/upsampling layer" is defined practically, depending on the image size, which improves the accuracy of classification. Additionally, the planes count is determined as the dividing quotient the cells count in the layer of input by two to the power of the doubled pairs count "convolutional - pooling/upsampling layer" to retain the total cells count in the layer after pooling/upsampling. This process halves the layer planes size in width and height, automating the structure definition of the model layers. The deep Boltzmann machine model offers several advantages over the traditional deep Boltzmann machine. These include preresizing the input image, determining the number of limited Boltzmann machines empirically to increase classification accuracy, and setting the neurons count in the hidden layers as double the neurons count in the visible layer to satisfy the Kolmogorov theorem on the representation of multidimensional continuous functions by a superposition of one-dimensional continuous functions. This model automates the definition of the model layer's architecture. The intelligent classification method of chicken eggs developmental state, based on deep associative neural networks, can be applied in intelligent systems for classifying the results of chicken eggs candling visualization during incubation in industrial poultry production.
When incubating eggs, it is crucial to consider several factors, such as maintaining the appropriate temperature and humidity, monitoring the composition of the ventilated air, and using high-quality, fresh, and fertilized eggs to ensure the production of healthy offspring. Before placing the eggs in the incubator, it is important to conduct ovoscopy to check for egg integrity and the possibility of development. Ovoscopy involves transilluminating the eggs to select high-quality ones without structural damage, which is essential for obtaining a healthy brood.
The ovoscopy process is the first stage before using the incubator. This process is usually repeated
2-3 times during the incubation period to check for defects in the shell and inside the eggs, the
absence of an air chamber, and the presence of any embryo abnormalities. This helps in identifying
and rejecting eggs with pathologies or other developmental disorders [
The cause of non-developing offspring may be due to abnormal egg shape, thinned or damaged
shell (pits, protrusions, roughness, dark spots), calcareous growths on the shell, the presence of
foreign objects or clots, the presence of two yolks at once, the yolk location not in the center or its
0000-0003-3841-7373 (E. Fedorov); 0000-0002-6614-4133 (T. Utkina); 0000-0002-3954-3796 (O. Nechyporenko);
0000-0002-0210-9582 (M. Leshchenko); 0000-0003-0000-6077 (K. Rudakov); 0000-0002-3318-3347 (I. Zubko)
displacement, the yolk remains in place when the egg is turned, displacement of the air chamber, or
the absence of the embryo [
To reduce the number of unsuitable eggs for incubation, the following conditions must be
observed from the very beginning: collect eggs 3-4 times a day; do not incubate eggs laid after 18
hours; ensure hygiene monitoring and check for the absence of mechanical damage to the protective
egg film when collecting with appropriate machines, automated lines, or by workers: to replace the
litter as needed. Chicken eggs should be stored for no more than 5 days, while duck and turkey eggs
should not be stored for more than 8 days, and goose eggs for no more than 10 days. Maintain the
air temperature within the range of 10-15 degrees Celsius and the relative humidity between 70-80%
during storage. Ensure that the temperature does not exceed +27 degrees Celsius, as this can lead to
the development of the embryo, resulting in unsuccessful incubation. Similarly, the temperature
should not fall below +8 degrees Celsius, as this could cause irreversible chemical changes in the
eggs [
Throughout the entire development cycle, the embryo changes color and condition. As a result, ovoscopy is used to monitor the proper development of the fetus. Ovoscoping is conducted in a warm, darkened room while observing fire safety regulations. It is crucial that the eggs are exposed to heat rays for no more than 15-20 seconds during ovoscopy to prevent overheating, which can render them unsuitable for incubation.
A fresh egg, when examined with an ovoscope, should display the following characteristics: a consistent shell; a small air pocket at the rounded end of the egg; the yolk positioned centrally or slightly closer to the rounded end, with indistinct boundaries; when the egg is rotated, the yolk should rotate with some resistance; and there should be no foreign objects inside the egg.
The following defects should be observed during conducting ovoscopy on non-viable egg
specimens [
Not all eggs selected based on external signs during candling will necessarily hatch, as only fertilized eggs will do so during the early stages of ovoscopy. Fertilized eggs can be identified after 5-7 days of incubation during the next stage of ovoscopy. There will be no signs of embryo development during candling.
Monitoring egg development allows for the assessment of the progress of the incubation process
and the identification of eggs with frozen fetuses due to hypothermia, overheating, or sticking to the
film for the timely removal of such specimens from the incubator. This will prevent the spread of
infection and reduce energy losses for incubating non-viable specimens, as well as adjust the
microclimate parameters in the incubator in order to increase the percentage of offspring
hatchability [
Today, to increase the efficiency of monitoring the state of development of the embryo of poultry
eggs during otoscopy, various methods of visualizing the incubation process are used: tomography,
magnetic resonance and infrared visualization, microscopy, ultrasound, thermal difference, digital
signal processing, etc. [
Intelligent identification methods are used to conduct a qualitative assessment of the results of otoscopy of poultry eggs. This allows for high classification accuracy, improved quality of monitoring of the incubation process, and reduced costs in industrial poultry production.
Currently, deep neural networks [
The first class of deep non-associative neural networks are convolutional networks such as:
1. LeNet-5 [
Attention), and a two-layer perceptron.
6. Compact Convolutional Transformer (CCT) [
Consequently, the issue of creating an effective deep associative neural network that addresses these concerns is relevant.
The first class of such networks are autoencoders such as: ⦁ ⦁
Convolutional autoencoder [
Variational autoencoder [
The second class of such networks is the deep Boltzmann machine [34]. In order to achieve our goal, we need to accomplish the following tasks:
Deep non-associative networks have the following drawbacks [
Difficulty in determining the parameters of the architecture of a deep associative neural network (patch size, size and count of layers, etc.); Insufficiently high training speed; Insufficiently high recognition accuracy.
To create a model to classify the state of chicken eggs using a convolutional autoencoder. To develop a model to classify the state of chicken eggs using a deep Boltzmann machine. To choose the quality criteria for the egg state classification method.
To determine the structure of the egg state classification method.
To perform a numerical research of the offered egg state classification approach.
The aim of the work is to improve the quality of classification of the state of chicken eggs by using deep associative neural networks. 2. Creating an ovoscopy model based on a convolutional autoencoder
Figure 1 shows a convolutional autoencoder for sample recovery, a dynamic non-recurrent network with a hierarchical architecture.
The input image in this type of convolutional autoencoder is pre-resized, and the “convolutional - pooling layer” pairs count is defined practically, based on the image size. Additionally, the planes count is determined automatically as the quotient of dividing the cells count in the layer of input by a power of two. This power is equal to twice the “convolutional - pooling/upsampling layer” pair count. This method allows for the preservation of the total cells count in the layer after pooling/upsampling, which effectively reduces/increases the layer planes size by two times in width and height.
In contrast to MLP, the input layer and output layer have the same neurons count and receive the same data. The hidden layers count is always odd, and the number of neurons in the hidden layers is fewer than the neurons count in the input/output layer. This decreases as we approach the central (code) hidden layer. Every autoencoder consists of an encoder and a decoder. When there's only one hidden layer and a linear activation function, the autoencoder becomes similar to PCANN.
The autoencoder implements an auto-associative memory (a pair of samples ( , ), = , representing its element (cell)) and restores (extracts) the stored sample by the key sample corresponding to the input vector . The most important property of a convolutional
Downsampling layer 1 autoencoder is that the same network with the same connection weights can store and reproduce several stored samples.
1 r i a p N r i a p
layer, and are the cell planes count, is the connection area of the layer plane and , ̑ are the convolutional layers count.
1. = 1. 2. The output signal for the convolutional layer is being calculated.
( , ) =
( ℎ ( , )), ∈ {1, . . . , }2, ∈ 1, ,
22 , ≤ ̑ /2, = {22( ̑ − +1), > ̑ /2, 1( ) + ∑ 1( , 1, ) ( + ), ∈ 1 where 1( , 1, ), ( , , ) is the connection weight, ( , ) is the cell output.
If ≤ ̑ /2, then to calculate the output signal for the downsampling layer (downsample by a factor of 2) where ( , )is the connection weight, ( , ) is the cell output.
If ≥ ̑ /2, then to calculate the output signal.
( , ) =
max { (2 + , )}, ∈{0,1}2 ∈ {1, . . . , }2, ∈ 1, , = 22 , (2 + , ) = {
̑ ( , ), = 2 ,
̑ ( , ) > 2 , ∈ {0,1}2, ∈ {1, . . . , }2,
∈ 1, , = 22( ̑ − ) where ( , ) is the cell output.
If ≤ ̑ , then increment , go to 2.
The output signal for the output convolutional layer is being calculated.
( , 1) = sigm( ℎ ( , 1)), ℎ ( , 1) = (1) + ∑ = 1 ∑ ∈ ̑ ( , , 1) ( + , ), (1) (2) (3) (4) (5) (6) where ( , , 1) is the connection weight, ( , 1) is the cell output. 3. Creating an ovoscopy model based on the deep Boltzmann machine
neurons
neurons
where is the probability, is the neural network energy increment.
The Deep Boltzmann Machine model is presented as follows.
During the bottom-up pass (recognition phase), the following steps are executed: 1-6. Bottom-up pass (recognition phase) (steps 1-6) 1. Initialization of binary vectors of hidden neurons states 1( ) = ( 1( ) , . . . , ( ) ( )), ∈ 1, . 2. = 1, (0) = .
Positive phase (step 3)
The state of hidden neurons of the layer ∈ 1, ( ) is being calculated. = { 1 + exp (− ( )( ) − ∑ ( −1)
=1 ( ) ( +1) ( +1))
1 + exp (− ( )( ) − ∑ ( −1)
=1 ( ) ( −1)) , < ,
= . 1, 0,
The state of visible or hidden neurons of the − 1 layer ∈ 1, ( −1)is being calculated. = 1 + exp (− ( −1)( ) − ∑ ( )
=1 ( )
= 1, , 1 < ≤ . ( −1) = { 1, 0,
The state of hidden neurons of the layer ∈ 1, ( ) is being calculated.
=
1 1 + exp (− ( )( ) − ∑ ( −1) ( −1)( )
=1 ( −1)) ( ) = { 1, 0,
with , 0, with 1 − , = , (8) (9) (10) (11) (12) (13) (14) = 1 + exp (− ( )( ) − ∑ ( +1) ( +1))
1 + exp (− ( )( ) − ∑ =1 ( +1) ( +1) ( +1))
1, 0,
The state of hidden neurons of the + 1 layer ∈ 1, ( +1)is being calculated.
Top-down traversal (spawning phase) (steps 7-11)
The state of hidden or visible neurons of the layer ∈ 1, ( ) is being calculated. 1 + exp (− ( +1)( ) − ∑ ( ) =1 { 1 + exp (− ( +1)( ) − ∑ ( ) =1 ( +1) ( ) − ∑ ( +2) =1 ( +2) ( +2)) 10. If > 0, then go to step 8. 11. The state of visible neurons of zero layer ∈ 1, (0) is being calculated.
( +1) = { 1, 0, (1) (1))
(0) = { (0) , . . . , 1, 0, , , , > 0,
= 0.
= − 1, , 0 < ≤ − 1.
(15) (16) (17) (18) (19) (20) (21)
The result is a pattern = ( 1 4. Selection of quality criteria for classifying the state of chicken eggs. The next criteria were elected to evaluate the training of the suggested mathematical models of deep associative neural networks:
Accuracy criterion: = 1 ∑ =1 [ = ̑ ] → max,
̑ = { 1, = arg max , 0, ≠ arg max .
Categorical cross-entropy criterion: 1
=1 =1
,
(22)
where – is -th model output vector, ∈ [
Development of a Training, Validation and Test Dataset Development of a Deep Association Neural Network Model
Training a Deep Association Neural Network Model Recognition by Deep Associative Neural Network Model
Accuracy>ε˄CCE<δ – +
W
Numerical research was carried out using the Chicken benchmark [37], which consisted of RGB images sized 1080x800. Out of the total 380 photos, 80% were randomly chosen for the learning set, while the remaining 20% were used for the test and validation sets. Since the proposed deep associative neural networks do not employ recurrent connections, their parameters were determined using a GPU. The “Keras” module was utilized to realise the suggested deep associative neural networks, and “Google Colab” was selected as the environment of software. to the beginning of the neural network.
Table 2 shows the structure of the deep Boltzmann machine model, with a resizing layer added to the beginning of the neural network.
Figure 5 shows the relationship between the categorical entropy and the “convolutional pooling/upsampling layer” pairs count for the convolutional autoencoder.
Figure 6 shows the impact of the iterations count on the loss (measured by categorical entropy) for the deep Boltzmann machine model.
Figure 7 shows the effect of the RBM count on the loss (measured by categorical entropy) for the deep Boltzmann machine model. According to the numerical research results, the following recommendations can be made: ⦁ ⦁ ⦁ ⦁
The minimum number of iterations for a convolutional autoencoder is 11 (Fig. 4). The optimal “convolutional – pooling/upsampling layer” pairs count for a convolutional autoencoder are 2 (Fig. 5).
The minimum iterations count for a deep Boltzmann machine is 17 (Fig. 6, Fig. 7).
The best RBM count for a deep Boltzmann machine is 3 (Fig. 7).
Different image classification methods were explored to enhance the accuracy of identifying the condition of chicken eggs; deep associative neural networks are the most effective method currently available.
The model of associative autoencoder offers several advantages compared to the traditional version. It pre-resizes the input image, determines the "convolutional - pooling/upsampling layer" pairs count empirically based on the image size to improve classification accuracy, and calculates the planes count by dividing the cells count in the layer of input by two to the power of the double the number of pairs, ensuring the total cells count in the layer remains constant after pooling/upsampling. This reduces the layer planes size by half in width and height, automating the process of determining the model's layer architecture. The modified deep Boltzmann machine model offers several advantages over the traditional model. First, the input image is pre-resized, which is beneficial for processing. Second, the number of limited Boltzmann machines is determined empirically, leading to improved classification accuracy. Third, the neurons count in the hidden layers is set to double the neurons count in the visible layer, in accordance with the Kolmogorov theorem. This automated approach simplifies the determination of the model's layer structure. The developed method uses deep associative neural networks for intelligent classification of the state of chicken eggs. This method can be applied in intelligent systems to classify the state of eggs.
The author(s) have not employed any Generative AI tools.