Abstract
In recent years, the rapid development of the Internet of Things (IoT) and AI deep learning technologies has enhanced communication between devices and increased the availability of image data. In smart agriculture, sensors remotely monitor crop information, and cameras detect leaf growth and health, thereby enhancing farm productivity and quality. However, most object detection algorithms, such as YOLOv4, have too many parameters for real-time deployment on edge computing devices. To address this, we propose the MobileNetv3-YOLOv4 (MobYOLv4) architecture, which uses MobileNetv3 as the backbone and depthwise separable convolution to reduce complexity. This approach balances accuracy and detection speed. Experimental results reveal that the lightest MobYOLv4 architecture reduces parameters by 82.31%, achieves 29.48 FPS (a 38.14% increase), and has an accuracy of 97.74%. The MobYOLv4 with widths 1.4 architecture, optimized with width multiplier, reduces parameters by 78.03%, achieves 28.87 FPS, and has an accuracy of 98.84%. The MobYOLv4 with widths 1.8 architecture achieves the highest accuracy at 99.65%, reduces parameters by 72.62%, and achieves 27.91 FPS. Furthermore, the MQTT IoT transmission protocol is used to manage farm data and facilitate device communication. The proposed adaptive width multiplier strategies in MobYOLv4 for lightweight object detection balance accuracy improvement per additional parameter with overall model accuracy, effectively optimizing the model. The width multiplier can significantly impact the model’s parameters and actual detection accuracy.
Abstract from the paper’s OpenAlex record, shown for indexing and research. Copyright remains with the authors or the publisher. Read at the source ↗
The text the model read, with the source record’s inline markup (italics, sub- and superscripts) rendered and its character entities decoded; quote offsets refer to the raw text.
Claims 10
Claim · model-written summary
The lightest MobYOLv4 architecture reduced model parameters by 82.31%.
Verbatim
e0“the lightest MobYOLv4 architecture reduces parameters by 82.31%”
Claim · model-written summary
The lightest MobYOLv4 architecture achieved 29.48 FPS, a 38.14% increase.
Verbatim
e0“the lightest MobYOLv4 architecture reduces parameters by 82.31%, achieves 29.48 FPS (a 38.14% increase)”
Claim · model-written summary
The lightest MobYOLv4 architecture attained 97.74% accuracy.
Verbatim
e0“has an accuracy of 97.74%”
Claim · model-written summary
MobYOLv4 with widths 1.4 reduced parameters by 78.03%.
Verbatim
e0“The MobYOLv4 with widths 1.4 architecture, optimized with width multiplier, reduces parameters by 78.03%”
Claim · model-written summary
MobYOLv4 with widths 1.4 achieved 28.87 FPS.
Verbatim
e0“achieves 28.87 FPS”
Claim · model-written summary
MobYOLv4 with widths 1.4 achieved 98.84% accuracy.
Verbatim
e0“has an accuracy of 98.84%”
Claim · model-written summary
MobYOLv4 with widths 1.8 achieved the highest accuracy at 99.65%.
Verbatim
e0“The MobYOLv4 with widths 1.8 architecture achieves the highest accuracy at 99.65%”
Claim · model-written summary
MobYOLv4 with widths 1.8 reduced parameters by 72.62%.
Verbatim
e0“reduces parameters by 72.62%”
Claim · model-written summary
MobYOLv4 with widths 1.8 achieved 27.91 FPS.
Verbatim
e0“achieves 27.91 FPS”
Claim · model-written summary
Adaptive width multiplier strategies balanced accuracy improvement per additional parameter with overall model accuracy.
Not reportedValue
Verbatim
e0“balance accuracy improvement per additional parameter with overall model accuracy”