An efficient strawberry segmentation model based on Mask R-CNN and TensorRT

Currently, artificial intelligence (AI), particularly computer vision (CV), has numerous applications in agriculture. In this field, the production and consumption of strawberries have experienced great growth in recent years, which makes meeting the growing demand a challenge that producers must face. However, one of the main problems regarding the cultivation of this fruit is the high cost and long picking times. In response, automatic harvesting has surged as an option to address this difficulty, and fruit instance segmentation plays a crucial role in these types of systems. Fruit segmentation is related to the identification and separation of individual fruits within a crop, allowing a more efficient and accurate harvesting process. Although deep learning (DL) techniques have shown potential for this activity, the complexity of the models leads to difficulty in their implementation in real-time systems. For this reason, a model capable of performing adequately in real-time, while also having good precision is of great interest. With this motivation, this work presents a efficient Mask R-CNN model to perform instance segmentation in strawberry fruits. The efficiency of the model is assessed considering the amount of frames per second (FPS) it can process, its size in megabytes (MB) and its mean average precision (mAP) value. Two approaches are provided: The first one consists on the training of the model using the Detectron2 library, while the second one focuses on the training of the model using the NVIDIA TAO Toolkit. In both cases, NVIDIA TensorRT is used to optimize the models. The results show that the best Mask R-CNN model, without optimization, has a performance of 83.45 mAP, 4 FPS, and 351 MB of size, which, after the TensorRT optimization, achieved 83.17 mAP, 25.46 FPS, and only 48.2 MB of size. It represents a suitable model for implementation in real-time systems.