A Combined Network for Tomato Leaf Disease Recognization Based on the Improved EfficientNet

Identifying tomato leaf disease with artificial intelligence algorithms is significant for global agriculture. However, the existing algorithms have the following problems: (1) Model parameters and calculations are too large. (2) Results affected by light intensity. (3) The introduction of modules a...

Full description

Saved in:
Bibliographic Details
Published in2022 IEEE International Conferences on Internet of Things (iThings) and IEEE Green Computing & Communications (GreenCom) and IEEE Cyber, Physical & Social Computing (CPSCom) and IEEE Smart Data (SmartData) and IEEE Congress on Cybermatics (Cybermatics) pp. 395 - 400
Main Authors Li, Xiao-Mei, Xu, Min, Li, Tian-Yu, Xu, Wei, Gai, Rong-Li, Hu, Ling-Yan, Wang, Zu-Min
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.08.2022
Subjects
Computational modeling
Convolution
DO-Conv
EfficientNet
Feature extraction
HSV color space
lightweight
Neural networks
Semantics
Social computing
tomato disease
Visualization
Online AccessGet full text

Cover

More Information
Summary:Identifying tomato leaf disease with artificial intelligence algorithms is significant for global agriculture. However, the existing algorithms have the following problems: (1) Model parameters and calculations are too large. (2) Results affected by light intensity. (3) The introduction of modules adds a large number of parameters. To solve the above problems, this study proposes a combined network model which composed of lightweight neural networks, HA-DSEfficientNet. Firstly, introduce the Depthwise Over-parameterized (DO) module to EfficientNetV2B0 network to effectively improved the model's learning capability without increasing the number of additional parameters. Secondly, simplify the Fused mobile inverted bottleneck convolution (Fused-MBConv) module to reduce the number of parameters and computation of the model. Finally, introduces the fusion of HSV features and RGB features to the disease recognition task and uses DSEfficientNet and AlexNet to extract RGB and HSV features of images, respectively. This approach reduces the effect of light intensity on recognition accuracy and enhances the features' semantics and discriminability. The final experimental recognization accuracy reached 99.24%, which was 2.02% higher than the accuracy of the network before improvement.
DOI:10.1109/iThings-GreenCom-CPSCom-SmartData-Cybermatics55523.2022.00089