Sistema automático de classificação de imagens térmicas para detecção de mastite subclínica bovina

Abstract

Brazil occupies a prominent position in the world dairy production sector. However, this sector faces an obstacle well known by producers, the bovine mastitis. It causes many production losses, and it is necessary to develop tools that enable its early detection, reducing the time, cost and subjectivity associated with the determination of the diagnosis of subclinical mastitis. Thus, the objective of this study was to develop a computational methodology capable of receiving digital thermal images, which allows segmentation and automatic classification, helping to diagnose bovine mastitis. The data survey was carried out in three milk production units, located in the municipalities of Capoeiras and Pesqueira, in the Mesoregion Agreste, Microregion of Ipojuca Valley, State of Pernambuco and in the Municipality of Russas, Ceará. To develop the automatic segmentation methodology, images from 24 animals were used, in different clinical conditions (healthy, with clinical and subclinical mastitis) determined according to the selection criteria. The thermal images of the udder of the animals were obtained by infrared thermographic camera, FLIR i60, obeying the left anterolateral, right anterolateral, posterior and inferior frames, four images per animal. For the development of the methodologies using sequential transfer learning, 600 images from the MammoTherm (human breast cancer) bank and 165 images of 360 x 360 pixels, referring to the database of 55 cattle, classified into distinct groups "Healthy" and "Subclinical Mastitis", were used. The automatic segmentation indicated representativeness of the segmented area of 19, 15, 37 and 36% of the total pixels for healthy animals [32.9 - 33.86 °C] ± 0.99. For the subclinical mastitis picture, the percentage representation ranged from 21.84 to 69.5% of total pixels [34.45 - 34.98 °C] ± 0.87. The representation of animals with clinical mastitis ranged from 78.5 to 85.85% [35.34 - 35.75 °C] ± 0.67. The algorithm for automatic segmentation allowed differentiating the images of healthy animals, with subclinical and clinical mastitis. The predictive model STL_bayesian_CBAM-ResNet50 achieved the best performance (97.28%) compared to the other models, 92.1% (STL_bayesian-ResNet50) and 88.03% (STL_ResNet50), respectively. The computational methodology applied to the study, from thermal images of the udder of dairy cows, contributed significantly to the automatic detection of healthy animals and animals with subclinical mastitis.