A novel lightweight model for tea disease classification based on feature reuse and channel focus attention mechanism
When collecting disease images in tea gardens, the collected images often include complex backgrounds. The complex backgrounds affect the recognition accuracy of traditional deep learning techniques. To improve the recognition accuracy, the traditional classic convolutional neural network (CNN) mode...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-01-01
|
| Series: | Engineering Science and Technology, an International Journal |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2215098624003264 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1850083934405132288 |
|---|---|
| author | Junjie Liang Renjie Liang Dongxia Wang |
| author_facet | Junjie Liang Renjie Liang Dongxia Wang |
| author_sort | Junjie Liang |
| collection | DOAJ |
| description | When collecting disease images in tea gardens, the collected images often include complex backgrounds. The complex backgrounds affect the recognition accuracy of traditional deep learning techniques. To improve the recognition accuracy, the traditional classic convolutional neural network (CNN) models require higher model complexity. This leads to a significant increase in computational cost and affects the running speed on edge computing devices. On the other hand, the popular vision transformer(ViT) model has a higher recognition accuracy as it has better global feature expression ability than CNN model. Since ViT models has a higher model complexity than CNN models, leading running slow on edge computing devices. In view of this, we propose a lightweight model named Lightweight Tea Diseases Detection Network (LTDDN). The LTDDN has higher recognition accuracy under the interference of complex backgrounds. Specifically, first, we propose the lightweight channel focus attention mechanism (CFA). The CFA focus the key features of the disease, improving the recognition accuracy. Second, we propose the feature reuse module (FRM). The FRM significantly reduces the parameters and computational costs of the model, making the model more lightweight. Finally, we propose a feature enhancement mobile inverted bottleneck convolution module(FEMBCM). The FEMBCM solves the problem of detail feature loss caused by traditional convolution downsampling. It improves recognition performance without increasing the complexity of the model. The analyses show that the parameter of LTDDN is only 1/30 of ResNet50 and 1/50 of ViTB16, which is more suitable for running in edge computing devices. Experimental results across self-built tea disease datasets, Mini-ImageNet, and PlantVillage datasets show LTDDN outperforms classic CNN, lightweight CNN, and popular ViT models in recognition accuracy. Based on the proposed model, an Android app has been developed to realize real-time, offline recognition of tea diseases in the field. |
| format | Article |
| id | doaj-art-5996402bdf694e4aac6a01f41e84b874 |
| institution | DOAJ |
| issn | 2215-0986 |
| language | English |
| publishDate | 2025-01-01 |
| publisher | Elsevier |
| record_format | Article |
| series | Engineering Science and Technology, an International Journal |
| spelling | doaj-art-5996402bdf694e4aac6a01f41e84b8742025-08-20T02:44:12ZengElsevierEngineering Science and Technology, an International Journal2215-09862025-01-016110194010.1016/j.jestch.2024.101940A novel lightweight model for tea disease classification based on feature reuse and channel focus attention mechanismJunjie Liang0Renjie Liang1Dongxia Wang2Yazhou Bay Innovation Institute, Hainan Tropical Ocean University, Sanya, 572022, China; School of Marine Information Engineering, Hainan Tropical Ocean University, Sanya, 572022, China; College of Mathematics and Computer Science, Zhejiang A&F University, Hangzhou, 311300, ChinaYazhou Bay Innovation Institute, Hainan Tropical Ocean University, Sanya, 572022, China; School of Marine Information Engineering, Hainan Tropical Ocean University, Sanya, 572022, China; Corresponding author.College of Mathematics and Computer Science, Shanxi Normal University, Taiyuan, 037000, ChinaWhen collecting disease images in tea gardens, the collected images often include complex backgrounds. The complex backgrounds affect the recognition accuracy of traditional deep learning techniques. To improve the recognition accuracy, the traditional classic convolutional neural network (CNN) models require higher model complexity. This leads to a significant increase in computational cost and affects the running speed on edge computing devices. On the other hand, the popular vision transformer(ViT) model has a higher recognition accuracy as it has better global feature expression ability than CNN model. Since ViT models has a higher model complexity than CNN models, leading running slow on edge computing devices. In view of this, we propose a lightweight model named Lightweight Tea Diseases Detection Network (LTDDN). The LTDDN has higher recognition accuracy under the interference of complex backgrounds. Specifically, first, we propose the lightweight channel focus attention mechanism (CFA). The CFA focus the key features of the disease, improving the recognition accuracy. Second, we propose the feature reuse module (FRM). The FRM significantly reduces the parameters and computational costs of the model, making the model more lightweight. Finally, we propose a feature enhancement mobile inverted bottleneck convolution module(FEMBCM). The FEMBCM solves the problem of detail feature loss caused by traditional convolution downsampling. It improves recognition performance without increasing the complexity of the model. The analyses show that the parameter of LTDDN is only 1/30 of ResNet50 and 1/50 of ViTB16, which is more suitable for running in edge computing devices. Experimental results across self-built tea disease datasets, Mini-ImageNet, and PlantVillage datasets show LTDDN outperforms classic CNN, lightweight CNN, and popular ViT models in recognition accuracy. Based on the proposed model, an Android app has been developed to realize real-time, offline recognition of tea diseases in the field.http://www.sciencedirect.com/science/article/pii/S2215098624003264LTDDNTea disease classificationChannel focused attention mechanismFeature reuse moduleDetail feature compensationComplex background interference |
| spellingShingle | Junjie Liang Renjie Liang Dongxia Wang A novel lightweight model for tea disease classification based on feature reuse and channel focus attention mechanism Engineering Science and Technology, an International Journal LTDDN Tea disease classification Channel focused attention mechanism Feature reuse module Detail feature compensation Complex background interference |
| title | A novel lightweight model for tea disease classification based on feature reuse and channel focus attention mechanism |
| title_full | A novel lightweight model for tea disease classification based on feature reuse and channel focus attention mechanism |
| title_fullStr | A novel lightweight model for tea disease classification based on feature reuse and channel focus attention mechanism |
| title_full_unstemmed | A novel lightweight model for tea disease classification based on feature reuse and channel focus attention mechanism |
| title_short | A novel lightweight model for tea disease classification based on feature reuse and channel focus attention mechanism |
| title_sort | novel lightweight model for tea disease classification based on feature reuse and channel focus attention mechanism |
| topic | LTDDN Tea disease classification Channel focused attention mechanism Feature reuse module Detail feature compensation Complex background interference |
| url | http://www.sciencedirect.com/science/article/pii/S2215098624003264 |
| work_keys_str_mv | AT junjieliang anovellightweightmodelforteadiseaseclassificationbasedonfeaturereuseandchannelfocusattentionmechanism AT renjieliang anovellightweightmodelforteadiseaseclassificationbasedonfeaturereuseandchannelfocusattentionmechanism AT dongxiawang anovellightweightmodelforteadiseaseclassificationbasedonfeaturereuseandchannelfocusattentionmechanism AT junjieliang novellightweightmodelforteadiseaseclassificationbasedonfeaturereuseandchannelfocusattentionmechanism AT renjieliang novellightweightmodelforteadiseaseclassificationbasedonfeaturereuseandchannelfocusattentionmechanism AT dongxiawang novellightweightmodelforteadiseaseclassificationbasedonfeaturereuseandchannelfocusattentionmechanism |