A fast and accurate deep learning method for strawberry instance segmentation

Abstract

Fruit instance segmentation is a key element for autonomous picking systems. This paper proposes a methodology for instance segmentation of strawberries using deep learning techniques. This methodology presents the following changes in relation to the well-known Mask R-CNN network: a new architecture is designed for the backbone and the mask network; the object classifier and the bounding-box regressor are removed; the nonmaximum suppression algorithm is replaced with a new region grouping and filtering algorithm without increasing the order of complexity. For this research, a new high-resolution data set of 3100 strawberry images is obtained, along with the corresponding manually-segmented ground-truth images (StrawDI_Db1 database). In addition to using widelyused metrics based on Average Precision (AP), a new performance metric is proposed: the Instance Intersection Over Union (I2 oU), to assess different options of instance segmentation techniques. This material, which is published and available to the research community, allows the rigorous performance comparison between future methodologies. The results obtained in the 200 images included in the test set show that the proposed methodology significantly reduces the inference time (10 fps vs. 5 fps) while maintaining competitive results to the original Mask R-CNN for mean AP (43.85 vs. 45.36) and mean I2 oU (87.27 vs. 87.70) metrics. Thus, the proposed method can be considered as the most promising one in view of its possible integration in automatic strawberry picking systems.

Existing System

SVM (Support Vector Machine)

Proposed System

The methodology of instance segmentation proposed in this article is based on the Mask R-CNN model. In this work, this network has been modified to significantly reduce the computational cost associated with the inference stage. This is a decisive factor since, on the one hand, the embedded systems where the network is to be implemented have limited computing resources and, on the other hand, their practical application requires high processing capacity.

Architecture

BLOCK DIAGRAM