Mobile-based Primate Image Recognition using CNN

Six out of 25 species of primates most endangered are in Indonesia. Six of these primates are namely Orangutan, Lutung, Bekantan, Tarsius tumpara, Kukang, and Simakobu. Three of the six primates live mostly on the island of Borneo. One form of preservation of primate treasures found in Kalimantan is by conducting studies on primate identification. In this study, an android app was developed using the CNN method to identify primate species in Kalimantan wetlands. CNN is used to extract spatial features from primate images to be very efficient for image identification problems. The data set used in this study is ImageNets, while the model used is MobileNets. The application was tested using two scenarios, namely using photos and video recordings. Photos were taken directly, then reduced to a resolution of 256 x 256. Then, videos were taken in approximately 10 to 30 seconds with two megapixel camera resolution. The results obtained was an average accuracy of 93.6% when using photos and 79% when using video recordings. After calculating the accuracy, the usability test using SUS was performed. Based on the SUS results, it is known that the application developed is feasible to use.


INTRODUCTION
Primates play an important role in research to improve human health. Despite being a small part of animals that are used in the biomedical field, primates share genetic and physiological similarities with humans, making them invaluable in developing treatments and vaccines for human diseases [1]- [3]. Primates also have important ecological functions in the ecosystems they inhabit, such as providing pollination services [4] and recognized as having an important role as a disperser of plant seeds [5]- [7] particularly in tropical rainforests [8].
Bristol Conservation and Science Foundation (BCSF), International Primatological Society (IPS), International Union for Conservation of Nature Species Survival Commission Primate Specialist Group (IUCN / SSC PSG), and Conservation International (CI) have published as many as 25 of the world's most endangered primate species [9]. Primates are threatened with extinction because of their natural ecosystems, mostly for commercial purposes, such as agriculture, livestock, commercial logging, timber harvesting, and fuelwood production [10]. Other factors are hunting, disease, climate change, loss, degradation and fragmentation of their habitats [11].
One form of preservation of primate treasures found in Kalimantan is by conducting studies on primate classification. With the primate classification, it is hoped that the public will be more familiar with protected primate animals. In artificial intelligence, especially machine learning, classification can be done by classifying primate images using a mathematical approach. Many methods can be used to classify images, including the Convolutional Neural Network (CNN). The CNN method can be used to solve problems related to image classification [17]- [20], recognition [21]- [24], and object detection [25], [26]. CNN is considered to have excellent accuracy and quality [17], [20], [27]- [29]. CNN uses a multi-layer neural network consisting of one or more convolutional blocks and sub-sampling layers, after which there are one or more fully connected layers and an output layer [30]. CNN's design is inspired by the structure of the mammalian visual system [31]. The most advantageous aspect of CNN is reducing the number of parameters in ANN [32]. Freytag conducted a study to identify chimpanzees using Log-Euclidean CNN with an accuracy of 92% [33]. Deb performed facial recognition of endangered primates, such as golden monkeys, lemurs, and chimpanzees, using the CNN method [34] with an accuracy of 88%. CNN is also used in the classification of medical images [35].
Based on the previous explanation, the researchers offer a solution for classifying primate images using the CNN method on an Android-based application. There are three types of primate images classified, namely Orangutan, and Bekantan. The application will be tested using two scenarios to calculate its accuracy. Also, usability testing was carried out using the System Usability Scale (SUS) to determine how easy it was for users to use the interface of the application being developed.

METHODS
The method used in this study consists of two stages, namely application development and application testing, as shown in Figure 1.

Application Development
Three steps were carried out at the development stage, namely interface design, code generation, and application functional testing.

1.1 Dataset
The dataset used in this development is ImageNet. There are two species of animals to be classified, namely Orangutan (Pongo pygmaeus) and Bekantan (Nasalis larvatus). Samples of images from each class can be seen in Figure 2

1.2 Design
At this stage, the interface design and the architecture design of primate image classification were carried out. The application interface consists of three pages, namely a splash screen, an object detection page, and a page to display the detection results or descriptions of the objects that have been identified. Image classification architecture design can be seen in Figure 3.  Figure 3, the first step in the primate identification process was to receive input in the form of photos or videos of primates directly. Photos are taken directly, then reduced to a resolution of 256 x 256. For videos, videos are taken for approximately 10 to 30 seconds with a 2-megapixel camera resolution. The second step was pre-processing to speed up the identification process. The kind of pre-processing that was done is to reduce the size of the photo to 256 x 256 pixels.
In the third step, the imaged primate images were compared using the MobileNets model to determine the primate's name. MobileNets is one of the Convolutional Neural Network (CNN) architectures that can be used to overcome the need for excessive computing resources. Once the name of the primate is known, the application will search for descriptions on the Wikipedia site based on the primate name. The final step, the output of the application was to display the name and description of the primate.

1.3 Coding
At this stage, code was generated for the application of interface design and image classification architecture in Android-based applications. The coding was made using android studio, firebase, and tensorflow [28]. The application was designed to run well on android 4.1 and up.

1.4 Functional testing
The functional testing was carried out using BlackBox testing by testing all functional features in the application, such as photo and video detection functions and the function of displaying results. It also ensures that there are no errors during the identification process of primate animals, resulting in the application not running properly or errors.

2 Application testing
There were two types of testing at the application testing stage: testing the accuracy and testing the application's usability using the System Usability Scale (SUS) [36]. Application accuracy testing was carried out to obtain the accuracy of primate recognition from the application being developed. Testing was carried out using two scenarios as follows: First Scenario: Photos of 3 types of animals were downloaded; each type consisted of 30 photos. The photos were recognized using an application that had been developed, then the results of the recognition were recorded to calculate accuracy.
Second Scenario: The application was tested by classifying using video, the application was highlighted to the primate video, then the recognition results were recorded to calculate accuracy.
After calculating the accuracy of the application, the next step was to test usability using SUS, which consists of 10 questions with a ring scale answer, from 1 to 5. Figure 1 shows that the user strongly disagrees with the test statement, and number 5 states that he strongly agrees with the test statement [37]. Usability testing is used to see how easy the user uses the application interface.

RESULTS AND DISCUSSION
This current study used 300 primate images for testing the first scenario and 30 videos for testing the second scenario. The purpose of this experiment is to measure the accuracy of the architecture that we design. After that, we also use SUS to measure how easy it is to use the interface of the developed application. Figure 3 shows the results of developing an android application using the CNN method. The user will first enter the primate identification interface to identify the desired object, as shown in Figure 4a.

CONCLUSIONS
Primate image classification is a difficult task to undertake because there are so many variations. In this paper, we try to use MobileNet to classify primate images. The experimental results show that MobileNet works very well, similar to the results obtained by previous researchers who also used MobileNet. The average accuracy in this study was 93.6% when using photos and 79% when using video recordings. Based on the usability test using SUS, 70 was obtained, which means that the application is feasible to use.