Lightweight Fruit-Detection Algorithm for Edge Computing Applications

Zhang, Wenli; Liu, Yuxin; Chen, Kaizhen; Li, Huibin; Duan, Yulin; Wu, Wenbin; Shi, Yun; Guo, Wei

doi:10.3389/fpls.2021.740936

Cited by 19 publications

(26 citation statements)

References 35 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This variation of weight size influences the tested real-time detection of model to support the claim of Lawal [ 8 ]. Apart from Mask-RCNN unable to meet the less than 50 ms standard of real-time detection proposed by Zhang et al [ 40 ], all YOLO-related models were able to achieve this standard as shown in Table 2 . YOLOv5-LiNet and YOLOv5-LiNetC having the same detection time of 2.6 ms is faster than 55.6 ms of Mask-RCNN, 2.8 ms of YOLOv5n, 3.4 ms of YOLOv5-Efficientlite, 3.5 ms of YOLOv5-MobileNetv3, 2.9 ms of YOLOv5-GhostNet, 2.9 ms of YOLOv5-LiNetBiFPN but slower than 2.4 ms of YOLOv5-LiNetFPN, 2.4 ms of YOLOv5-ShuffleNetv2 and 2.2 ms of YOLOv4-tiny.…”

Section: Resultsmentioning

confidence: 91%

YOLOv5-LiNet: A lightweight network for fruits instance segmentation

Lawal

2023

PLoS ONE

View full text Add to dashboard Cite

To meet the goals of computer vision-based understanding of images adopted in agriculture for improved fruit production, it is expected of a recognition model to be robust against complex and changeable environment, fast, accurate and lightweight for a low power computing platform deployment. For this reason, a lightweight YOLOv5-LiNet model for fruit instance segmentation to strengthen fruit detection was proposed based on the modified YOLOv5n. The model included Stem, Shuffle_Block, ResNet and SPPF as backbone network, PANet as neck network, and EIoU loss function to enhance detection performance. YOLOv5-LiNet was compared to YOLOv5n, YOLOv5-GhostNet, YOLOv5-MobileNetv3, YOLOv5-LiNetBiFPN, YOLOv5-LiNetC, YOLOv5-LiNet, YOLOv5-LiNetFPN, YOLOv5-Efficientlite, YOLOv4-tiny and YOLOv5-ShuffleNetv2 lightweight model including Mask-RCNN. The obtained results show that YOLOv5-LiNet having the box accuracy of 0.893, instance segmentation accuracy of 0.885, weight size of 3.0 MB and real-time detection of 2.6 ms combined together outperformed other lightweight models. Therefore, the YOLOv5-LiNet model is robust, accurate, fast, applicable to low power computing devices and extendable to other agricultural products for instance segmentation.

show abstract

Section: Resultsmentioning

confidence: 91%

YOLOv5-LiNet: A lightweight network for fruits instance segmentation

Lawal

2023

PLoS ONE

View full text Add to dashboard Cite

show abstract

“…The results from Tab.24 and 25 illustrate that RTNet attached with ResNet-18/CSPNet, which are widely used backbones in CV [76], [77], [78], [79], [80], and self-attention, which is widely used module in NLP [81], [82], [83], share similar and promising performances. These indicate that RTNet not only has splendid adaptability but also has great potential to couple with SOTA architectures in other AI fields.…”

Section: E4 Ablation Studymentioning

confidence: 99%

Respecting Time Series Properties Makes Deep Time Series Forecasting Perfect

Shen¹,

Wei²,

Wang³

2022

Preprint

View full text Add to dashboard Cite

How to handle time features shall be the core question of any time series forecasting model. Ironically, it is often ignored or misunderstood by deep-learning based models, even those baselines which are state-of-the-art. This behavior makes their inefficient, untenable and unstable. In this paper, we rigorously analyze three prevalent but deficient/unfounded deep time series forecasting mechanisms or methods from the view of time series properties, including normalization methods, multivariate forecasting and input sequence length. Corresponding corollaries and solutions are given on both empirical and theoretical basis. We thereby propose a novel time series forecasting network, i.e. RTNet, on the basis of aforementioned analysis. It is general enough to be combined with both supervised and self-supervised forecasting format. Thanks to the core idea of respecting time series properties, no matter in which forecasting format, RTNet shows obviously superior forecasting performances compared with dozens of other SOTA time series forecasting baselines in three real-world benchmark datasets. By and large, it even occupies less time complexity and memory usage while acquiring better forecasting accuracy. The source code is available at https://github.com/OrigamiSL/RTNet.

show abstract

“…To reduce the computational complexity and improve the speed of deep learning architectures number of models has been proposed such as Mobile Net, MobileNetv2, ShuffleNetv1 and ShuffleNetv2, Yolo tiny series etc., As fruit size in the actual environment is variable, these networks are unable to predict small size fruits with small number of layers. A Light model with edge devices Light-CSP Net has been proposed 158 to overcome this issue. In the structure of Light-CSP Net down sampling layer with cross scale fusion replaced the maximum pooling layer and FPN is replaced with deep shallow layer.…”

Section: Fruit Feature Extraction Using Deep Learning Approachesmentioning

confidence: 99%

Selective fruit harvesting: Research, trends and developments towards fruit detection and localization – A review

Kumar

Mohan

2022

Proceedings of the Institution of Mechanical Engineers, Part C:

View full text Add to dashboard Cite

Progressive application of multidisciplinary research and development pushes the evolution of automation in many subsectors of agriculture to increase productivity, economic growth and environmental preservation with the help of robotics and artificial intelligence. Fruit harvesting robots have been developed mainly to provide support in the field for limited labour resources, to enable selective harvesting, to improve the efficiency and to preserve the quality of fruits. Even a small delay in harvesting can cause a maximum impact to the quality of the fruit. Selective fruit harvesting is an integration of different subcomponents. This paper provides a brief analysis of the techniques in selective fruit harvesting for the past 6 years starting from 2017 to 2022, associated principles, limitations and directions for future challenges. The first subcomponent is the vision system, it captures the information about the fruit in a tree canopy to perform efficient 2D and 3D localization. Hence getting accurate information from the vision system is more essential even in a complex agricultural environment. The detection of fruit from the vision system is normally done with two major methods such as traditional image processing and deep learning approaches. The result shows that the traditional methods provide high efficiency but the colour similarity, complex backgrounds and lightning conditions often makes failure in detection. Shortage of standard dataset and high-powered processing devices hinders the development of deep learning algorithms. Also, the usage of large data sets reduces the training speed even in pre-trained networks. For fruit grasping and detachment, the detection of plucking point is more essential to preserve the quality and for further storage. The elaborate description about the framework, limitations in current sensing, fruit and picking point detection algorithms provides guidelines to the researchers in building a fully automated robotic system to increase the processing speed and production rate.

show abstract

Lightweight Fruit-Detection Algorithm for Edge Computing Applications

Cited by 19 publications

References 35 publications

YOLOv5-LiNet: A lightweight network for fruits instance segmentation

YOLOv5-LiNet: A lightweight network for fruits instance segmentation

Respecting Time Series Properties Makes Deep Time Series Forecasting Perfect

Selective fruit harvesting: Research, trends and developments towards fruit detection and localization – A review

Contact Info

Product

Resources

About