Desarrollo de un modelo de reconocimiento de ganado bovino usando deep learning

Abstract

Ecuador is one of the countries where it has a great economic impact on the livestock sector, many of the owners have large amounts of paddocks, several of these have opted to use technology to increase their productivity and avoid the loss of their animals, due to the infrequent control of their animals, since this demands time and greater physical effort from their employees, leaving unnoticed the different activities that must be carried out day after day in the other paddocks. At present, the use of artificial intelligence in the area of cattle farming has become an advantageous opportunity, optimizing its material and human resources, so that its arduous hours of work have now been converted into minutes, thus obtaining very good results and increasing your earnings. This research focused on the identification of cattle regardless of their breed, class, color and size, by using machine learning and deep learning, which are more efficient in processing large amounts of data to perform a recognition model. objects, taking into account factors such as climate, time and image resolution. In the implementation of the neural network learning, Yolov4 was used, which uses an object detection algorithm called darknet, and this architecture provides an adjustable base to adapt it for personal use according to the learning needs. This project was oriented towards the creation of a model that allows cattle to be recognized without considering their race or color, through a single class called "bovine". For the collection of aerial photographs, a DJI air 2 model drone was used. achieving a total of 1630 images, as each photo session has been generated in different terrains, these have been divided into 4 data sets, allowing the development of 8 training sessions combining and uniting each dataset. The hardware used in the training was a 10th generation Intel Core i7 10750H six-core processor, 64 GB of DDR4 RAM, an NVIDIA RTX 2060 graphics card with 6 GB GDDR6 of video memory, along with the operating systems used Ubuntu 24.04 LTS and Windows. 10Enterprise 21H2. The tools used were the CUDA 10.1 and cuDNN 7.6.5 technologies for the Yolo v4 application, in addition to the Python programming language together with its OpenCV and Matplotlib libraries, the LabelImg labeling tool for label generation and Jupyter Notebook as a notebook. execution of code statements. The best model obtained from each training was selected for its evaluation in detection, two types of tests were carried out, one for training and validation, and another for performance. In the evaluation test, the values of precision, sensitivity, reference value and the average precision obtained from the results during training were obtained. In the performance test, a video taken from a drone from above was used and a manual evaluation of the detection results of each model was carried out. To obtain the most optimal model, an evaluation based on a Likert scale was used to apply a point rating according to the results obtained in the two tests applied. The results of the training model 1 were satisfactory since it was the one that had the best rating and presented the best results with different sources of test videos.