Clasificador de inventario por medio de una red neuronal convolucional con sincronización a una base de datos

Abstract

In this project, the development of a procedure based on the algorithm is carried out Faster RCNN ResNet 101 for classifying an inventory and recording it in a database, in addition to being able to compare the results obtained, adatabase portion for post-validation. The Faster RCNN ResNet 101 algorithm has been implemented in detection problems of objects, obtaining good results with models of low complexity and medium complexity, which is why there has been a growing interest in applying this algorith in detection and classification of objects. Furthermore, no application of the algorithm is presented in Store inventory problems, which is why it is the reason for this project. To carry out the classification of products, it is necessary to carry out the collection of a database, taking into account the division for training, validation and post-validation. This database is obtained by means of two professional cameras by labeling the images with Labelimg and processing the images to increase the information provided by the database. The input parameters of the object detection system are then selected. With base 70% of the database, the training and validation of the model is carried out, for through convolutional networks and networks of relationship proposals, through the model Pretrained Faster RCNN ResNet 101. A variation of an own parameter "learning coefficient" is carried out to find the classifier model. The results show that it is possible to perform a detection of objects through the pretrained model with an approximate performance of 96.2%, but it is necessary to make a database with possible human failures, such as the greater amount of noise in order to obtain an efficient model, since the performance of the model has a direct relationship with the noises in the database. Regarding the counting of the products, a module was generated at the output of the object detection, which shows the number of products per class with an efficiency of the 88.6%, Finally the data was uploaded to FireBase, a dynamic database with various connection tools on platforms.