Influência da quantização de redes neurais convolucionais na classificação de imagens de abelhas portadoras de pólen

Abstract

The recognition of pollen-carrying bees can provide important information about the operational conditions of beehives. It can also indicate the intensity of pollination activity, which is a fundamental aspect for many plant species and of great commercial interest to agriculture. In this context, embedded systems are commonly used in the process of acquiring and processing images of bees at the entrances of beehives. However, these systems generally have constraints regarding computational capacity and storage space. In this work, the performance of fourteen deep convolutional neural network models in solving the problem of classifying pollen-bearing and non-pollen-bearing bees from a set of images was analyzed. The study also analyzed how the quantization process influences these results. Quantization allows for a reduction in inference time and model size by performing calculations and storing numbers in a lower-precision format, which can be advantageous in contexts based on the use of embedded systems. The experimental results indicate that it is possible to reduce the inference time and/or the model size without diminishing the performance metrics of accuracy, precision, recall, F1-score, and false positive rate, facilitating deployment in an embedded system. The best result was achieved by the AlexNet network quantized to int8, reducing the inference time and model size by 45.63% and 71.73%, respectively.