The present paper aims to study image processing algorithms to accelerate and facilitate the evaluation of the harvest condition in tomato farms. In order to achieve this, two different deep learning models are trained and combined with counting methods to produce a harvest monitoring system for embedded applications using an Intel^® Movidius^TM and an affordable RGB camera. The first model detects the location of cherry tomato clusters, while the second estimates the fruit’s maturity. The results are compared to a baseline implementation based on segmentation. Next, a multiple counting method based on regions of interest is applied to the detected clusters in videos to count the tomatoes at different maturity stages. In order to produce a more robust counting, a tracking system is implemented which uses temporal information to find the unique tomato clusters in videos. In the evaluation stage, the obtained location results indicate an intersection over union (\( IoU \)) of about \(89\%\) when using the MobileNetV1 as a feature extractor and choosing the appropriate location anchors. The maturity estimation results indicate better performance for the trained algorithm as compared to the baseline, providing a root mean square error of \(7.7\%\). The best results were obtained when combining the fully learned solution with the tracking system, correctly counting the majority of the tomato clusters at multiple maturity stages.

本文旨在研究图像处理算法，以加速和促进对番茄农场收获条件的评估。为了实现这一点，两个不同的深度学习模式进行培训，并具有计数方法来生产使用英特尔嵌入式应用的收获监控系统相结合^® Movidius ^TM和负担得起的RGB相机。第一个模型检测樱桃番茄簇的位置，而第二个模型估计水果的成熟度。将结果与基于细分的基线实现进行比较。接下来，将基于关注区域的多重计数方法应用于视频中检测到的簇，以对不同成熟阶段的西红柿进行计数。为了产生更可靠的计数，实​​施了一个跟踪系统，该系统使用时间信息在视频中查找独特的番茄簇。在评估阶段，所获得的位置结果表明交集的交集（\（IoU \））约为\（89 \％\）将MobileNetV1用作特征提取器并选择适当的位置锚点时。成熟度估计结果表明，与基线相比，训练算法的性能更好，提供的均方根误差为\（7.7 \％\）。将完全学习的解决方案与跟踪系统结合使用，正确计算多个成熟阶段的大部分番茄簇时，可以获得最佳结果。