In this paper guinea pig classi cation using deep learning imaging methods was performed on the Nvidia DIGITS 6. Models capable of distinguishing skinny, abyssinian and crested fur types were created in the process. To increase the classi cation accuracy empty images (with only the background) were added to the data set. Upon evaluation, the created model recognized the animals correctly from images taken in various household backgrounds.
Robotic systems are nowadays increasingly appearing in various industries. This
trend is also represented in various animal care facilities like farms, daries,
shelters [
In this work, deep learning techniques were utilized to create a classi cation model of guinea pigs in di erent home environments (living room, o ce, corridor, etc.), to explore the possibilities of bringing such systems into the world of household animals. Creating such a model was important, to understand how machine learning algorithms would adapt to live creatures, while keeping a high accuracy of the prediction and short inference times needed in robotics.
To address these problems GoogLeNet was used. The pre-trained model
connects various techniques known from Deep Learning like convolutions, pooling,
adding softmax and more [
Gathered results, will perform as a base for future projects of animal social and care systems. Example appliances could include: { automatic feeding and cleaning systems, { automatic pet door management, { illness and status detectors, { or mobile, home robots allowing the owners to check on their pets using mobile software. 1.1
Related works about animal classi cation were published in regard of wild animal
monitoring [
Similarly convolutional neural networks were used to recognize 20 species
common in North America [
Another approach was taken for classifying di erent animals for automated
species recognition [
All mentioned authors follow a similar pattern when using machine learning for image classi cation, but none of them concern household animals. All presented projects face di erent problems from those, that could be encountered in a safe, indoor environment. Hence, di erent data acquisition techniques had to be used. 1.2
DIGITS (Nvidia Deep Learning GPU Training System) was used in this project.
It is an open-source project for training deep neural networks (DNNs). The
software simpli es common deep learning tasks such as managing data, designing
and training neural networks and monitoring performance in real time. [
In this paper GoogLeNet was used as the model of choice.
Deep Learning (DL) is a machine learning technique growing in popularity over
the past few years. It is connected to the fact that it becomes more useful than
before thanks to the amount of available training data and advances in computer
hardware/software [
To understand how DL works, we should rst de ne what learning actually
is. A simple de nition was provided by Mitchell [
This can apply to di erent kind of tasks, performance measures and experiences. In this paper we will focus on the task of object classi cation using computer vision.
Classi cation of objects is based on describing to what category a given input
belongs. This can be used in robotics for tasks like delivering foods and drinks
to clients by the Willow Garage PR2 robot [
This paper focused on supervised learning, which means that all data set items were associated with a label (each guinea pig image was added to a speci c category). In practice, it means that the algorithm knew how to classify certain objects with similary properties from the start. 2.1
Articial neural networks are a subgroup of algirthms that are used for machine learning. The main idea behind them is creating arti cial neurons, wchich are implemented as a non-linear function over a linear combination of input features. Each neuron generally consist of one to multiple inputs with wages, activation function, bias and one output.
In neural network algorithms we can tweak several parameters, that will greately impact the nal output: number of epochs, learning rate, solver type and many more.
{ Epoch - one complete pass of the data in the data set. The amount of epochs
should be determined through tests. Small number of epochs often leads to
bad predictions, while a big number leads to over tting.
{ Learning rate - describes the rate at which the network abandons old
beliefs, for new ones to take their place. The value must be correct, values
that are too big or too small may lead to bad predictions.
{ Solver type - contains information about how the weights are updated for
the network. This project used Stochastic Gradient Descent (SGD) [
When working with machine learning algorithms, we can encounter two problems, that are the result of our actions - under tting and over tting.
Over tting takes place, when the model is training the data too well. That happens when noise, details or random uctuations are taken into consideration, which negatively impacts the performance of the model on any new data. In such case, the parameters should be adjusted to constrain the amount of detail that the model learns.
Under tting on the other hand is usually the result of big learning rates. The model becomes too general, which in the end gives bad results for object classi cation. To provide a solution to the problem, usually adjusting parameters or using di erent ML algorithms should be used. 2.2
Stochastic Gradient Descent (SGD) is one of the most popular algorithms in DL. It is an extension to the rst-order optimization algorithm: Gradient Descent. In SGD, the gradient is an expectation, which may be estimated even using a small set of samples. SGD has proven to work very well with deep learning models. While it doesn't guarantee nding the local minimum, it usually nds a very low value of the cost funcion quickly.
The estimate from the example x minibatch m0 can be written as follows: g = m10 O Pim=01 L xi; yi; , where the loss is L (x; y; ) = log p (yjx; ) for each example.
Adam is an optimization algorithm that is used for iteratively updating the network weights based on the training data. The algorithm combines two extensions of the SGD: Adaptive Gradient Algorithm (AdaGrad) and Root Mean Square Propagation(RMSProp). Instead of adapting the parameter learning rates based on the average rst moment (the mean) as in RMSProp, Adam also makes use of the average of the second moments of the gradients (the uncentered variance). 2.3
Convolutional Neural Networks (CNNs) are used for data, that has a known grid-like topology. The name comes from the fact, that it employs an operation called convolution, which is a kind of linear operation (instead of general matrix multiplication).
Convolutions are operations on two functions of a real value argument. The
convolution can be represented as s(t) = (x w)(t), where x is a single input, w
is a valid probability density function and t is time. It leverages three important
ideas that can improve machine learning systems: sparse interactions, parameter
sharing and equivariant representations [
Each layer of CNNs consists of three distinguishable stages: producing sets of linear activations, detector stage and pooling. Pooling is a method that instead of giving the output of the neural net, provides a summary statistic of all nearby outputs. This is especially helpful, if images of variable size are given as an input. Moreover it helps with feature extraction (as it is known, neural networks loose data over time, with each layer) from convolutional layers. 2.4
GoogLeNet was introduced at ILSVRC 2014 competition, where it took rst place with a result of 6.67% error rate (which is close to human level performance). The architecture consisted of 22 layers (27 with pooling) of the Deep CNN reducing the number of parameters to 4 million (60 million compared from AlexNet).
The main innovation between normal CNNs and GoogLeNet was the implementation of Inception modules. The modules ran several small convolutions in order to reduce the number of parameters. Moreover batch normalization, RMSprop and image distortions were used. Data from the previous layer was run over four 1x1 convolutions, one 3x3 convolution and one 5x5 convolution, with a 3x3 pooling added simultaneously. You can see the network presented on the graph on Fig. 1.
The Inception module 3x3 and 5x5 condolutions ratio increases as higher
layers are achieved. This is due to the fact, that stacking Inception modules on
top of each other produces an e ect where as features of higher abstraction are
captured by higher layers, their spatial concentration is expected to decrease [
The highest pro of the network is high inference speeds. GoogLeNet was designed to be computational e cient, so that it could be run on devices with limited computational power or low-memory footprint, making it a good choice for robotics and its applications. 2.5
While Deep Learning is performing well for many cases of image classi cation, it still has disadvantages, that should be considered when picking the right method. The most known cons of the method are: { small data sets can produce bad results, { long calculation time is a big factor, { debugging is extremly hard, { picking good neural net parameters takes practice, { it's hard to gather the logic behind results.
With recent advantages in Deep Learning some of the factors are less limiting. The learning duration for example, is being decreased using parallellization. The data was collected by recording a square video of each guinea pig over a period of 30 seconds in di erent environments and then extracting frames as an image. Each frame was then lowered in resolution to 256x256 px in order to t GoogLeNet requirements. Example images taken from the data set can be seen in Fig. 2.
The entire data set consisted of 1098 images. From the initial data set - 25% (274) images were excluded for model validation and 10% (110) were excluded to calculate the test data loss and accuracy. Moreover 32 photos in di erent environments (animal cage, sleepingroom, guestroom, balcony and bathroom) were taken after the training to test the model behaviour. The visual representation of the training data set can be seen on Fig. 3.
To ensure good accuracy of the model, images for each guinea pig had to cover the whole anatomy of the animal. To achieve that, the following camera positions were covered: { facing the mouth, { both sides front and back facing, { rear view of the animal, { top view with di erent distances to the guinea pig.
Moreover empty images (without guinea pigs) were added with the same room backgrounds where previous photos were taken, to increase classi cation accuracy in distinguishing wanted objects. This set contained 191 images mixed with the guinea pig data set.
Guinea pigs phenotype highly depends on their breed. Therefore three di erent subjects of diverese ages were used in the experiment:
Using the data, four labels were produced and assigned to the following neural network classes, which provided a base for further classi cation: { None - when there is no guinea pig in the image, { Abyssinian, Crested and Skinny - fur type.
Finally the data was processed by the neural network, on di erent settings. Two parameters were changed during testing: learning rate and optimizer used, while the epoch count remained at 15. The settings can be seen below: 1. First test: { Epoch count: 15, { Learning rate: 0.001 with xed policy, { Optimizer: Stochastic Gradient Descent. 2. Second test: { Epoch count: 15, { Learning rate: 0.001 with step down policy, { Optimizer: Adam.
{ Epoch count: 15, { Learning rate: 0.01 with step down policy, { Optimizer: Stochastic Gradient Descent. 4
Final results for classi cation models were gathered from the neural networks described in the previous section.
Performing the rst test gave good results, although over tting was discovered around epoch 13, with accuracy of 98,95% and loss of 0.04. Later epochs drastically fell in value, producing an accuracy of 65,63% and loss of 1.03. Overtting appeared because of the low learning rate from the very start, which should have been avoided.
Second test, using Adam as the solver type, provided the best results. The nal acurracy and loss after 15 epochs were 99.31% and 0.04 respectfully. A di erent training rate (0.01) was also tested, but it didn't provide any useful results.
The nal test gave good results, but not satisfactory - it produced an accuracy of 87,15% and loss of 0.32. That was not enough to be used for the guinea pig classi cation system (the predictions would give false-positives).
The nal classi cation model that was selected for further use and testing, was taken from test number two, using the Adam optimizer. Over 15 epochs with a learning rate of 0.001 ran on the GoogLeNet model on Nvidia DIGITS, it has provided the best results. Using more epochs and a di erent learning rate was tested afterwards, but it led to over tting of the data. You can see the nal result on Fig. 4. 4.1
After the model was created, a series of tests were performed to check, if the inference is correct. The rst set of tests consisted of images from the previously gathered data set. The prediction was always right for provided images, with the value between 70%-95%. You can see an example result on Fig. 5.
Moreover, as soon as acquiring enough information from predictions was done, tests on new images were performed (di erent environments, same guinea pigs) giving satisfactory results - guinea pigs were classi ed correctly on each provided picture containing the animal. Example result one such case can be seen on Fig. 6.
One failed prediction was encountered, when a picture of a background with a cat was used. The cat was badly classi ed as an abyssinian guinea pig. This happened due to the fact, that the data set didn't contain images that were in any case similar to the cat picture used. More cases like that can be produced with the provided model, as it was trained on speci c data in the rst place. Modern robotics highly depend on sensor readings of the surrounding environment. They often use camera input as one of the parameters to perceive the world. Due to that, imaging methods for decision-making were introduced.
In this paper Deep Learning was implemented for guinea pig classi cation in order to explore the possibilities of introducing household animal care using robotics and automation, while keeping them safe.
The provided GoogLeNet model from Nvidia DIGITS has proven successful in identifying the guinea pigs in di erent environments, even when taking images that weren't originally added to the data set. Some errors were observed (falsepositives) when no guinea pigs were present in the tested image. The model will behave poorly when other animals are present in the pictures, since the classi cation was based purely on guinea pigs.
Increasing the accuracy of the model, can greately improve the robot-animal interactions, allowing to tailor behaviours to speci c beings. This could be achieved by using modifying the learning rate, using more images, creating more labels or nally using a di erent optimizer or pre-built model. There are many possible variables to take into consideration, when building a model for a de ned task.
The guinea pig classi cation model after building with GoogLeNet, was able to provide results almost instantly. This is important, if used for robotics, because while waiting for an action, the environment can change quite drasticly. Moreover, such model are built to be deployed on an autonomous platform (Raspberry, Jetson TX2 or any other), so the memory usage will be limited, increasing the inference calculation time.
The created model proves that guinea pig fur recognition for robotic systems is possible. The project gave good results - the created model recognized the animals correctly from images taken in various household backgrounds. The prediction was acquired fast making the inference time low. This is especially important for robotic systems that deal with live animals, because the reaction times need to be rapid.
Future work might include robotic systems that monitor the state of speci c animals, adjust food distribution depending on image readings or alert when the guinea pig su ers from any kind of illness.
Moreover, di erent types of models can be employed to see, which one ts the needs the most. The project shouldn't be limited to GoogLeNet.