Snake identification is crucial in quickly and efectively treating snake bites. With over 2.7 million snake envenomings happening yearly, medical personal are in desperate need of tools that will ease their work as well as save patient lives faster. SnakeCLEF 2021 challenge, which is part of the LifeCLEF laboratory, has exactly the aforementioned goal. This paper presents our team participation at SnakeCLEF 2021. We developed 3 models based on CNN: GoogLeNet, VGG16 and ResNet-18 and ranked 5th with an F1-Country score of 0.785 using ResNet-18.
Human expansion and the amplification of animal habitat destruction is resulting in more and more contact with wildlife in urban areas. Snakes are one species afected by this phenomena, if previously they lived in forests, swamps or even deserts, they are now forced to crawl in urban areas in search of food and shelter. The negative impact is felt not only in animals but also humans, more and more are we in contact with venomous snakes which leads to deadly scenarios that put us on a knife-edge. Being able to quickly identify the species of snakes that came in contact with people will not only give medical personnel precious time but also will give us a better understanding of certain species of snakes and their mobility in this new human habitat.
Annually, according to WHO [
Manual identification is no easy feat, there are more than 3,500 species of snakes, 600 of which are venomous. Training doctors on each and every species is an impossible task, it would be not only time consuming but also very costly. Recent years have brought an exponential growth in A.I. research, of which image recognition seems to be the most benefited. Advances and expansion of the global smartphone market combined with high Internet Penetration Rate in low-income and middle-income countries has lead to an information boom. People now have access to suficient computing power at the tips of their fingers making it possible to classify in real-time almost anything. Snake identification and classification will improve the quality of lives beyond high-income countries and significantly improve epidemiology data and treatment outcomes.
The SnakeCLEF 2021 challenge [
This paper describes the participation of team "UAIC-AI", from the Faculty of Computer Science, “Alexandru Ioan Cuza” University of Iasi, Romania, at SnakeCLEF 2021 where we ranked 5th with an F1-Country of 0.785. The remaining of this paper is organized as follows: Section 2 describes state-of-the-art methods in snake identification, Section 3 details the model we developed and the submitted runs and then Section 4 details the results we obtained, finally Section 5 concludes this paper and presents future work.
SnakeCLEF 2020 [
Last year’s best result however was obtain by "Gokuleloop" team [
The developed methods, that address the SnakeCLEF 2021 challenge, are all based on
convolutional neural networks. We used 3 models, GoogLeNet [
In order to have a clearer picture of the competition, it is mandatory to compare this year’s dataset with the previous one. The 2020 dataset was much smaller in comparison to 2021, 245,185 training images were provided split into 783 species comparing to this year where there are now 414,424 images and 772 species. This means that in 2020 there were approximately 313 images per species and in 2021 there are now 536 images, a two-fold increase in images which will nonetheless result in a higher accuracy of the models. Additional geographical metadata (country and continent) for the image is also provided.
Analyzing the dataset yields important information related to the collection. Figure 1 shows age variation between snakes whilst Figure 2 illustrates geographic variations, demonstrating that the dataset has a variety of scenarios.
GoogLeNet is a system that has shown to be powerful in many novel applications. In [
Our goal was to experiment with this method and build an architecture that has filters with multiple sizes that can operate on the same level. We kept the starting filters from GoogLeNet which used a (1 x 1) convolution because training is time-consuming and this method was used to compute reductions before the expensive (3 x 3), (5 x 5) convolutions and max-pooling.
The GoogLeNet architecture is 22 layers deep (and 5 pooling layers). These layers are grouped in 9 inception modules, and each of these is connected to the average pooling layer.
The GoogLeNet model used was implemented in PyTorch [14] and trained using Nvidia CUDA on GPU, a massive help was also Google Colab [15] which saved us a lot of time by training on their machines. We used a Cross Entropy Loss and Adam optimizer [16].
We started the training process with a 0.001 learning rate and batch size of 64. Learning rate was adjusted at each epoch (in total we had 10 epochs) until we concluded that 0.001 is the best rate. We also enabled "aux_logits" which adds two auxiliary branches that can improve training. 3.3. VGG16 VGG16 is a well known model and although it is an older approach we wanted to compare it to the others as see how it performs, as in previous works [17] it showed potential. It is a 16 layer deep convolutional neural network with no residual blocks which improves on other models by replacing large kernel-sized filters with multiple 3 ×3 kernel-sized filters one after another.
We used the model implemented in Keras that had the pre-trained weights from ImageNet-1k. The model implemented uses Adam optimizer and the hyperparameters were tuned experimentally. Initially the learning rate was set at a base value of 0.001 then initially decreased to 0.0001. The learning decay function used was ReduceLROnPlateau with a factor of 0.6 and patience of 4 while monitoring the validation accuracy. The batch size set was 16 as this gave us a good compromise between training speed and resource consumption.
In order to prevent overfitting we used a Dropout Regularization [ 18] strategy during the training phase. In an efort to reduce loss, a dropout layer with a probability of 0.5 was added between the last 2 dense layers of the VGG16 model.
ResNet is the third and best model used. A study done in [19] outlines the performance overhead ResNet has over GoogLeNet, there is an almost 15% accuracy diference between the two, in favor of ResNet. The novelty and the performance of the method has lead us to experimenting with the model in order to test it in relationship with other neural networks. As other architectures often face the vanishing gradient problem, ResNet came with a solution called "identity shortcut connection” that skips one or more layers and acts like a much shallower network.
For this task a ResNet18 (from PyTorch) architecture pre-trained on ImageNet-1k [17] was ifne-tuned. We decided to keep the pre-trained weights only for the first three layers and freeze them. The final fully connected layer was reset and reshaped to 772 nodes so that it match the current number of classes in the dataset.
The model was trained with 14 epochs with batches of 32 and a learning rate of 0.0002. To prevent overfitting a dropout regularization strategy was chosen during the training phase. We also used Adam Optimizer and cross entropy loss. To reduce loss, inside the basic blocks of ResNet, a dropout with a probability of 0.5 was applied.
Additional information available for each image regarding the country were taken into account in the model. Table 1 illustrate a dataframe fed to the network in the training phase. We included the species name, country, continent, genus and family but also the image path (that is not presented in the table).
In this section we will discuss the results of the algorithms. Table 2 shows the results of our models. It is clearly seen that the best submission was "uaic_ai_submission_8" with an F1-Country of 0.785 according to the organizers, this ranked us 5th.
Submissions information: - uaic_ai_submission1: This submission was computed using a model with ResNet18 architecture resembling the training techniques from section 3.4 without taking into account the additional country information.
- uaic_ai_submission2: This submission was computed with GoogLeNet, consisted by all the information from the 3.2 subsection.
- uaic_ai_submission8: This submission was computed using a model with ResNet18 architecture resembling the training techniques from the 3.4 section and also taking into account the additional country information.
The best submission is represented by ResNet18 with country information. Further analysis into the GoogLeNet and VGG16 have revealed that due to aging architecture their performance is limited in comparison to more modern approaches such as ResNet. Scarce hardware resources and limited time have lead to slow progress in increasing ResNet performance and this will be the object of future work. Another research direction aims to address the imbalance of species in the dataset as we noticed that some species have more images than others. We would like to use augmentations from the Albumentation Library [20] for species that have few images in the dataset. The idea is simple, modifications are being done to dataset picture, like change of contrast, saturation, hue or brightness, in order to increase the dataset. Another technique we would like to use is MixUp augmentation [21], the technique is simple and they describe it as follows "mixing up the features and their corresponding labels. Neural networks are prone to memorizing corrupt labels. MixUp relaxes this by combining diferent features with one another (same happens for the labels too) so that a network does not get overconfident about the relationship between the features and their labels" [21].
In conclusion, this paper is focused on team UAIC_AI’s participation at SnakeCLEF 2021. We had good results with 3 submissions, the best one ranking 5th with an F1-Country of 0.785. For future work we would like to try much novel image algorithms as well as improve the score of the current methods.
Special thanks goes to A2 team. This work was supported by project REVERT (taRgeted thErapy for adVanced colorEctal canceR paTients), Grant Agreement number: 848098, H2020-SC1-BHC2018-2020/H2020-SC1-2019-Two-Stage-RTD. using improved googlenet for sparse-view ct reconstruction, Scientific reports 8 (2018) 1–9. [14] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, S. Chintala, Pytorch: An imperative style, high-performance deep learning library, in: H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, R. Garnett (Eds.), Advances in Neural Information Processing Systems 32, Curran Associates, Inc., 2019, pp. 8024–8035. URL: http://papers.neurips.cc/paper/ 9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf. [15] E. Bisong, Google colaboratory, in: Building Machine Learning and Deep Learning Models on Google Cloud Platform, Springer, 2019, pp. 59–64. [16] D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, 2017. arXiv:1412.6980. [17] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al., Imagenet large scale visual recognition challenge, International journal of computer vision 115 (2015) 211–252. [18] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, R. Salakhutdinov, Dropout: A simple way to prevent neural networks from overfitting, Journal of Machine Learning Research 15 (2014) 1929–1958. URL: http://jmlr.org/papers/v15/srivastava14a.html. [19] R. U. Khan, X. Zhang, R. Kumar, Analysis of resnet and googlenet models for malware detection, Journal of Computer Virology and Hacking Techniques 15 (2019) 29–37. [20] A. Buslaev, V. I. Iglovikov, E. Khvedchenya, A. Parinov, M. Druzhinin, A. A. Kalinin, Albumentations: Fast and flexible image augmentations, Information 11 (2020). URL: https://www.mdpi.com/2078-2489/11/2/125. doi:10.3390/info11020125. [21] F. Chollet, et al., Keras, https://keras.io, 2015.