Search

Wearing face mask in public has become a health protocol standard during this pandemic to prevent further spread of COVID-19. Even though the detection of inappropriate use of face mask is important considering that people sometimes ignore the health protocols by lowering their face mask so it does not cover their nose, studies regarding automatic detection of proper use of face mask are still few. Therefore, in this research we propose a multi-class image classification for detecting the proper use of face mask based on MobileNetV2 architecture as the base model. We also propose a trainable head model for the network, consisting of a depthwise convolution layer and two fully-connected layers, that gives high classification performance. The experimental results show that the proposed system gives a high multi-class classification performance with an accuracy of 97%, precision of 97%, recall of 97%, and F1-score of 97%. The running time of the proposed method is 265.94 seconds which is considered efficient compared with other models. Because of its light-weight network architecture, the proposed method is suitable for further implementation towards a real-time application of surveillance systems. Therefore, in this research we present the results of an initial experiment of the proposed model on a real-time detection system by using a web camera.

Nitrogen (N) is one of the main nutrients for plant growth, as a major component of chlorophyll and proteins closely related to leaf color, growth, and crop yield. The efficiency of nitrogen fertilization is needed to increase rice productivity. Adequacy of N nutrient can be observed in the leaf’s color. The use of Leaf Color Chart (LCC) is one way to analyze the need for nitrogen fertilizer in the rice field. This research proposed the development tool to determine the greenish scale of rice leaves automatically and to give advice on the appropriate and optimal dose of nitrogen fertilizer to the rice field. In this study, Arduino Uno is functioned as a Processing Unit and integrated with TCS3200 color sensor as an input component to detect the greenish level of rice leaf color. By using this tool, the prediction of the needs of nitrogen in the rice field can be determined more effectively, not depending on one’s perception of the leaf’s color scale or different light intensity. The measurements show that the system has an average scale misidentification of 2%, therefore this system can be used as an alternative to determining the dose of nitrogen fertilizer according to the conditions of the plants and the desired harvest target can be achieved optimally.

The agricultural sector really needs an application that able to estimate the effect of fertilization on plant growth patterns. The paper proposed the three dimensional (3D) simulation plant growth’s model of Glycine Max (L) Merrill/soybean plant using machine learning Multi-Layered Perceptron (MLP) method combine with Polynomial-Lindenmayer (Poly-L) system. The modeling parameters are the trunk/branches growth (L), the leaves width (W), and the number of branching (B) as the function of changes Nitrogen (N), Phosphate (P), and Potassium (K) elements in the fertilization process. The L, W, and B are modeled as the function of N, P, and K input using MLP method. Then, L, W, and B output are used as a variable to visualize plant growth into a 3D plant’s structure using the Poly-L System interpretation. The polynomial equation is used as a weighted factor according to the iteration of the L-System routine. The experimental results show that the MLP method is quite adaptable to the various changes of N, P, and K values and able to estimate the L, W, and B output. The average error of the trunk's growth prediction is 3.63%, the average error of leaf's width prediction is 3.72%, and the average error on the prediction of the branching's growth is 4.27%. The final result proved that the change of N, P, and K composition influenced the Poly-L System frames. Overall, the system has been running as expected.

Melanoma is the most aggressive of all skin cancers and its incidence has reached epidemic proportions. It is important to distinguish between benign and malignant melanoma as early as possible to increase the chance of recovery. The development of computational technology, especially machine learning and computer vision, made it possible to classify diseases based on their image. Detection of a disease by using image is beneficial because it can be done more easily, cheaply, quickly, and non-invasively than by using biopsy. The use of conventional machine learning and computer vision method makes their classification performance highly affected by the segmentation result of the skin lesion and the features selected for the classification process. The recent development of deep learning algorithm, such as CNN (Convolutional Neural Network), makes it possible to classify images without going through the process of image segmentation and manual features determination and give high performance with enough training data. Therefore, in this research we propose a deep convolutional neural network (CNN) to classify melanoma images into benign and malignant class. The proposed network architecture consists of several sets of convolutional layers and max-pooling layers, followed by a drop out layer and a fully-connected layer. From the experimental results on 352 test images, the proposed network gives the accuracy, sensitivity, and specificity of 84.76%, 91.97%, and 78.71%. The good performance of the built model hopefully can be developed for real application that can assist the expert to make better diagnosis and treatment.

The two-wheeled electric skateboard (TWS) is designed for a personal vehicle. A Fuzzy-PID control strategy is designed and implemented for controlling its motion. Basically, motions control of the TWS is performed by balancing the pitch position of the TWS. Performance of the designed controller is demonstrated experimentally. The Fuzzy algorithm updates the PID gains and therefore it can handle the changing of the TWS load. Contribution of Fuzzy-PID in reducing the electric energy consumption, which is an important issue in electrical system, is also evaluated. The Fuzzy-PID successes to reduce the electric energy consumption of the TWS compared to the conventional PID.

In this paper, we propose to classify supraspinatus muscle sEMG feature data to be contraction and relaxation data type. We use lightweight computational process feature extraction using moving average of absolute value and moving sum of absolute value. We use Neural Network with two hidden layer to classify sEMG feature. Two sEMG feature data are extracted successfully. Neural Network can classify input feature into contraction and relaxation data type with 99.8% accuracy. 99.8% accuracy are obtained with 1100 epoch of first experiment, and 577 epoch of second experiment.

Pencitraan tulang menggunakan ultrasound adalah sebuah teknik pencitraan yang aman karena bebas radiasi dan non-invasive . Namun, mendeteksi dan menentukan lokasi tulang untuk menemukan region of interest (RoI) adalah pekerjaan yang tidak mudah karena sifat citra b-mode yang memiliki derau yang tinggi dan reverberation artifacts . Kualitas citra dipengaruhi oleh pengguna dan batas antar area jaringan tidak jelas sehingga menyulitkan interpretasi. Dalam makalah ini, pendekatan deep learning menggunakan Region Proposal Networks diaplikasikan untuk mendeteksi RoI tulang dalam citra b-mode secara otomatis. Model arsitektur Faster Region-based Convolutional Neural Network disesuaikan agar dapat mendeteksi RoI tulang. Untuk mengevaluasi hasil, eksperimen secara in-vivo dilakukan menggunakan spesimen lengan manusia. Sebanyak 1.066 citra b-mode tulang dari enam subjek yang berbeda digunakan dalam fase latih dan fase pengujian. Metode yang diusulkan berhasil menentukan RoI tulang dengan baik dengan nilai mAP, akurasi deteksi, dan akurasi penempatan RoI masing-masing sebesar sebesar 0,87, 98,33%, dan 95.99%.

Pendeteksian patah tulang dengan X-ray atau CTscan menghasilkan gambar yang akurat tetapi memiliki efek negatif radiasi yang berbahaya. Makalah ini memaparkan penggunaan gelombang ultrasonik (US) sebagai alternatif pengganti kedua instrumen tersebut. Makalah ini menggunakan tulang femur sapi dan ayam dalam kondisi dengan dan tanpa daging, dengan patahan dibuat secara manual dengan pola patah transverse dan oblique . Pemindaian probe US menghasilkan citra B–mode dua dimensi. Pendeteksian tulang patah dilakukan menggunakan lima variasi desain arsitektur Convolutional Neural Network (CNN), yaitu CNN1–CNN5. Hasil uji coba menunjukkan bahwa desain arsitektur CNN4 memberikan hasil pengenalan kontur tulang dan klasifikasi tulang patah yang paling bagus dibandingkan desain arsitektur lain yang diuji, dengan akurasi 95,3%, sensitivitas 95%, dan specificity 96%. Hasil perbandingan dengan metode klasifikasi Support Vector Machine (SVM) dan k–Neural Network (k-NN) menunjukkan bahwa CNN memiliki unjuk kerja yang lebih unggul baik dalam hal akurasi, sensitivitas, maupun specificity .

Based on thermal acoustic data from the body tissue in upper arm that has been produced through thermal acoustic tomography method, classification system for the data has been built as a support of decision making about physiological abnormality. The advantages of the system built in this research is able to detect physiological abnormalities in the body tissues without the need for surgery (non-invasive method), such as to detect the existence of tumor or cancer and the damage of kidney or liver. Two classification methods has been evaluated for comparison, which are Backpropagation Neural Network (BP-NN) and Support Vector Machine (SVM). Using SVM that based on boundary optimization and BP-NN that based on error minimization as comparison, it is expected to obtain the most suitable classification method for decision making about physiological abnormalities in body tissues. The experimental results on 60 train data and 40 test data show that the accuracy for classification using Backpropagation Neural Network and Support Vector Machine (SVM) are 70% and 67.5%, respectively. In general, it can be proved that classification using Backpropagation Neural Network takes a longer running time (around 0.9 seconds) than classification using SVM (around 0.05 seconds).

Tomography is a method that can reconstruct the internal image of a non-invasively object, based on the measurements. On acoustic tomography, measured data is obtained from the output of ultrasonic transducer that is passed to the object. Velocity of waves that pass through the object is processed to obtain the impulse response and sampling projection. Once processed by the FFT, both these quantities is filtered by the back projection filter to obtain image reconstruction. This method that conducted to the air, water, mud, aluminum, and the combination of those materials, will showed gradations of color in the image in the form of a mesh, contour, pcolor, and surface. This shows the velocity difference when the ultrasonic wave reflected or transmitted by the material of measured object that can be indicated as the difference in the substance of the material in the measured object.