Nowadays, Closed Circuit Television (CCTV) monitoring systems, access control, and many other security-related applications incorporate facial recognition techniques. This disruptive tool differs from other biometric techniques, as faces can be recognized remotely. Thus, these applications can be incorporated in different institutions with the purpose of restricting access to unauthorized people, avoiding damages and losses to the property. The objective of this study is to identify people in controlled and uncontrolled environments within a university building that has suffered insecurity problems on several occasions. For this reason, it is used a convolutional neural network (CNN) architecture that was implemented on a GPU. The procedure consists of 3 phases: (i) training; (ii) learning and; (iii) classification. The CNN training stage was performed using the VGGFace2 dataset. The learning and generalization of deeply discriminatory facial embeddings of 512 bytes per face by joint supervision of softmax loss and center loss signals. The classification used a support vector machine (SVM) in several experiments with different amounts of classes. The efficiency of this approach was tested in real time on a sample of 128 students by means of quantitative metrics derived by confusion matrix. The overall accuracy in controlled and uncontrolled contexts was 96% and 71.43%, respectively. The proposed method establishes a starting point for the development of a more robust system in production environments.
This work presents a comparative study of prototype selection (PS) algorithms. Such a study is done over data-from-sensor acquired by an embedded system. Particularly, five flexometers are used as sensors, which are located inside a glove aimed to read sign language. Measures were taken to quantify the balance between classification performance and reduction training set data (QCR) with k neighbors equal to 3 and 1 to force the classifier (kNN) to the maximum. Two tests were used: (a)the QCR performance and (b) the embedded system decision in real proves. As result the Random Mutation Hill Climbing (RMHC) algorithm is considered the best option to choose in this data type with removed instances at 87% and classification performance at 82% in software tests, also the classifier kNN must be with k=3 to improve the classification performance. In a real situation, with the algorithm implemented. The system makes correct decisions at 81% with 5 persons doing sign language in real time.
The present work presents an analysis of the neighborhood criterion for the prototype selection (PS) in supervised machine learning classification algorithms. To do this, we use the condensed neighbor algorithm CNN to eliminate redundant data with the normalization of the distance to the centroid of each data subset criterion. This is done, in order to obtain the training matrix of the most optimal model. A selection of neighborhood criterion has been created from the quantification of the balance between the performance of the classification and the reduction of data set (CER). As proof of the test, we performed: (i) CER and (ii) real-time tests with the implementation of the algorithm within the WSN. The result is a data reduction of up to 88 % and a performance of the kNN classifier of 75%. It is concluded that the criterion of neighborhood with normalized distance must be less than or equal to 0.2 and the implementation of kNN with k = 1 obtains the best CER.
This work shows the use of Big Data and Data Mining techniques on vegetable crops data from a greenhouse by implementing the first version of a software tool, so called GreenFarm-DM. Such a tool is aimed at analyzing the factors that influence the growth of the crops, and determine a predictive model of soil moisture. Within a greenhouse, the variables that affect crop growth are: relative humidity, soil moisture, ambient temperature, and levels of illumination and CO2. These parameters are essential for photosynthesis, i.e. during processes where plants acquire the most nutrients, and therefore, if performing a good control on these parameters, plants might grow healthier and produce better fruits. The process of analysis of such factors in a data mining context requires designing an analysis system and establishing an objective variable to be predicted by the system. In this case, in order to optimize water resource expenditure, soil moisture has been chosen as the target variable. The proposed analysis system is developed in a user interface implemented in Java and NetBeans IDE 8.2, and consists mainly of two stages. One of them is the classification through algorithm C4.5 (chosen for the first trial), which uses a decision tree based on the data entropy, and allows to visualize the results graphically. The second main stage is the prediction, in which, from the classification results obtained in the previous stage, the target variable is predicted from information of a new set of data. In other words, the interface builds a predictive model to determine the behavior of soil moisture.
The present wireless sensor network system shows a data analysis approach within greenhouses in short cycle crops. This research, on the one hand, is carried out to reduce water consumption and improve the product by predicting the right moment of the irrigation cycle through the evapotranspiration criterion. On the other hand, an efficient electronic system is designed under the electronic standard. To define the best model to define the next irrigation in the crops in base to ground humidity, the algorithms are compared for continuous and discontinuous multivariate regressions. The results are evaluated with different criteria of prediction errors. As a result, the linear regression with Support Vector Machine model is chosen for counting an average deviation error of 7.89% and an error variability of 4.48%. In addition, water consumption is reduced by 20%, achieving better quality products.