Retinanet_G2S: a multi-scale feature fusion-based network for fruit detection of punna navel oranges in complex field environments

Abstract

In the natural environment, the detection and recognition process of Punna navel orange fruit using machine vision systems is affected by many factors, such as complex background, uneven light illumination, occlusions of branches and leaves and large variations in fruit size. To solve these problems of low accuracy in fruit detection and poor robustness of the detection algorithm in the field conditions, a new object detection algorithm, named Retinanet_G2S, was proposed in this paper based on the modified Retinanet network. The images of Punna navel orange were collected with Microsoft Kinect V2 in the uncontrolled environment. Firstly, a new Res2Net-GF network was designed to replace the section of feature extraction in the original Retinanet, which can potentially improve the learning ability of target features of the trunk network. Secondly, a multi-scale cross-regional feature fusion grids network was designed to replace the feature pyramid network module in the original Retinanet, which could enhance the ability of feature information fusion among different scales of the feature pyramid. Finally, the original border regression localization method in Retinanet network was optimized based on the accurate boundary box regression algorithm. The study results showed that, compared with the original Retinanet network, Retinanet_G2S improved mAP, mAP50, mAP75, mAP_S, mAP_M and mAP_L by 3.8%, 1.7%, 5.8%, 2.4%, 2.1% and 5.5%, respectively. Moreover, compared with 7 types of classic object detection models, including SSD, YOLOv3, CenterNet, CornerNet, FCOS, Faster-RCNN and Retinanet, the average increase in mAP of Retinanet_G2S was 9.11%. Overall, Retinanet_G2S showed a promising optimization effect, particularly for the detection of small targets and overlapping fruits.