The beekeeping is the art of cultivating the bees in the aim to remove from this industry the maximum performance with the minimum expenditure. The apiculture products marketed are the honey, wax, pollen, propolis and royal jelly. This activity of topping up contributes to the development of the livestock and to the protection of the Environment. This paper presents the application of deep convolutional neural network for pollen grains recognition based on their images classi cation. The neural network contains 8 hidden layers where rst 5 are convolutionnal neurones responsible for image representations and next 3 are fully connected layers for image classi cation. The obtained results proved the e ciency of the proposed approach for pollen grains recognition.
Honey is a nutritious food, with economic importance for many countries
worldwide. Nowadays, increasing attention was dedicated to the determination of the
geographical and botanical origin of the honey in the aim to de ne the di
erentiation character of honeys of di erent sources with a standard of quality and
authenticity competitive in the market [
The melissopalynology or palynology is applied to all knowledge in all
relations of any order that exist between the bee and the plant. As such it integrates
ecological, ethological and physiological research because it allows the use of
pollen grain as a biological marker in the vast context of the relations plant-bee.
Pollen analysis of honey is used to di erentiate the oral source used by bees,
the harvest period and the geo-climatic conditions of the concerned regions [
Computationally Motivated Biology is a eld that consist of studying
biology for modelling the biological systems using the computer science. For this
order, researchers study the behavior of a biological system then create tasks
as an arti cial model to facilitate the task for human beings. It is typically the
simulation of a natural phenomenon[
Like all machine learning techniques, deep learning aims to cause a system to
solve situations without calculating all the necessary parameters to solve them by
the implementer. The goal is to train a variable-parameter algorithm (a "black
box") to make a correct decision about a given task. Learning is done by
optimizing variable parameters to improve the decisions made. Deep learning techniques
have made great progress in many areas that range from image recognition [
Founding an e cient automatic classi cation system for pollen identi cation becomes a challenge that needs powerful techniques.This paper presents the application of convolutionnal neural network for pollen grains classi cation. The organisation of the paper was given as follow: Section 2 cites some works on the classi cation of honey pollen, while section 3 gives an idea about the implementation and results obtained during the experimentations and nally section 4 discusses the major conclusions. 2
Nowadays, automatic classi cation for pollen identi cation becomes a highly
active research eld. [
The year 2006 was the beginning of the deep learning, then, it has emerged as a new area of machine learning research. Since that, researchers focused on developing deep learning based techniques that impact signal and information processing, especially image processing that get the most part of deep learning development.
Deep Convolutional Activation Feature for Generic Visual Recognition(decaf)
is a python framework developped by Donahue et al. in [
Pollen23E is a set of 805 images divided in 23 species (classes). Each class
comprises of 35 images captured by a digital Bresser LCD microscope at a 40x
magni cation from di erent angles. Then the obtained images were transfered
to a laptop and segmented using the CorelDRAW1 software. [
In the following, we discuss the obtained results, in which, we based on the best training accuracy, best validation accuracy, best cross entropy and test accuracy to evaluate the e ect of number of iterations on pollen grains classi cation:
As seen in table 1, number of iteration a ected the obtained results in a manner that if we add more iterations, neural network builds better models. And this is clear in terms of rst three measures: { in terms of Training accuracy: This measure is the accuracy of applying the model on the training data, it is used to evaluate the model during backpropagation steps in order to improve the model. In table 1, we cited the best training accuracy obtained, as seen, the model is perfect since it correctly classi ed 99% to 100% of training data. Figure 1 shows a detailed updates of training accuracy during each iteration where (a), (b), (c), (d), (e), (f), (g) and (h) present 100, 200, 500, 1000, 1500, 2000, 2500 and 4000 iterations respectively. For all cases, we see in gure 1 how the training accuracy begins with small values (about 5%), then it improved iteration after iteration to reach 100%. { in terms of Cross Entropy: When we use cross entropy loss while training neural networks, we actually calculate the score function every time when compute gradients for the weights in the network. So, the objective is minimizing this measure, as seen in table 1, when we added more iteration, the neural network minimized the cross entropy which means we got better models. Figure 2 shows a detailed updates of cross entropy during each iteration where (a), (b), (c), (d), (e), (f), (g) and (h) present 100, 200, 500, 1000, 1500, 2000, 2500 and 4000 iterations respectively.
Fig. 2. Cross Entropy updates during each step For all cases, we see in gure 2 how the entropy begins with high values (about 5), then it improved iteration after iteration to converged to 0. While detailing the gures, we see that more we added iterations, the cross entropy became lower which means we have reduced information loss by improving the model. { in terms of Validation accuracy: This measure is the accuracy of applying the model on the validation data, as training accuracy, it is also used to evaluate the model during backpropagation steps in order to improve the model. In table 1, we cited the best validation accuracy obtained, as seen, the model is perfect since it correctly classi ed 95% to 100% of training data. Figure 3 shows a detailed updates of validation accuracy during each iteration where (a), (b), (c), (d), (e), (f), (g) and (h) present 100, 200, 500, 1000, 1500, 2000, 2500 and 4000 iterations respectively. For all cases, we see in gure 3 how the validation accuracy begins with small values (about 5%), then it improved iteration after iteration to reach 100%. Also, since validation accuracy is lower than training accuracy, we validated that our model did not su er from under tting problem. { in terms of Test accuracy: This measure is the accuracy of applying the nal model on the test data, it is used to evaluate the prediction of new images. Figure 4 shows the comparison of test accuracy according to number of iteration. In our case ( gure 4, the model recognized 76.6% of test images when we built a model in 100 iterations, while it was improved when we augmented the number of iterations to 200 iterations to recognize 85.1% of pollen grains, then it became xed despite the model has been improved based on previous measures. 3.3
To validate better our study, we compared our obtained results with other results
from literature. [
As seen, after reimplementing approaches proposed by [
Another comparison was done with same work, since they used deep neural networks as our objective, so we compared the time complexity taken by each approach with same conditions (number of iteration), Figure 6 shows the comparison:
As seen in gure 6, we reduced the training time, and the di erence of time reduced became higher when we increased the number of iteration for each approach, in which we gained from 0.47 minute up to 1.38, when we used 100 iteration, while the use of 4000 iteration proved that we can gain between 1.23 up to 2.17 minutes, and these values were obtained when we trained neural networks using 790 images, so we believe that our proposed approach will be more e ective in big data image analysis in terms of training time. 4
In this work, we applied a deep convolutional neural network for pollen identi cation from their images. The approach was divided into two processes, the rst process was feature extraction using ReLU non linearities convolutional calculation, where 5 layers were responsible for it, then, the dimentionality of those obtained features were reduced using pooling method to get vectors of size of 2048 element i which each vector represent an image in the dataset. After getting the vectors, they were used a s input to 3 fully connected layers for classi cation. The evaluation was based on four measures: Training accuracy, validation accuracy, cross entropy and test accuracy. As seen in the paper, the obtained measures' values proved that deep convolutional neural networks can be used as a good solution to automate the pollen grains classi cation. Also, we saw that the proposed approach avoid the under tting, this is proved by the validation accuracy that was lower than training accuracy.
These results motivated us for future works, we planned to develop more deep learning approaches, and combine it with metaheuristics.