Shunted Self-Attention via Multi-Scale Token Aggregation

Ren, Sucheng; Zhou, Daohong; He, Shengfeng; Feng, Jiashi; Wang, Xinchao

doi:10.48550/arxiv.2111.15193

Cited by 6 publications

(6 citation statements)

References 29 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Here, the local context enhancement term LCE(V) is introduced [23]. The function LCE(•) uses deep convolutional parameterization with a convolutional kernel size of 5.…”

Section: Bi-level Routing Spatial Attention Modulementioning

confidence: 99%

Detection of Orchard Apples Using Improved YOLOv5s-GBR Model

Sun,

Zheng,

et al. 2024

Agronomy

View full text Add to dashboard Cite

The key technology of automated apple harvesting is detecting apples quickly and accurately. The traditional detection methods of apple detection are often slow and inaccurate in unstructured orchards. Therefore, this article proposes an improved YOLOv5s-GBR model for orchard apple detection under complex natural conditions. First, the researchers collected photos of apples in their natural environments from different angles; then, we enhanced the dataset by changing the brightness, rotating the images, and adding noise. In the YOLOv5s network, the following modules were introduced to improve its performance: First, the YOLOv5s model’s backbone network was swapped out for the GhostNetV2 module. The goal of this improvement was to lessen the computational burden on the YOLOv5s algorithm while increasing the detection speed. Second, the bi-level routing spatial attention module (BRSAM), which combines spatial attention (SA) with bi-level routing attention (BRA), was used in this study. By strengthening the model’s capacity to extract important characteristics from the target, its generality and robustness were enhanced. Lastly, this research replaced the original bounding box loss function with a repulsion loss function to detect overlapping targets. This model performs better in detection, especially in situations involving occluded and overlapping targets. According to the test results, the YOLOv5s-GBR model improved the average precision by 4.1% and recall by 4.0% compared to those of the original YOLOv5s model, with an impressive detection accuracy of 98.20% at a frame rate of only 101.2 fps. The improved algorithm increases the recognition accuracy by 12.7%, 10.6%, 5.9%, 2.7%, 1.9%, 0.8%, 2.6%, and 5.3% compared to those of YOLOv5-lite-s, YOLOv5-lite-e, yolov4-tiny, YOLOv5m, YOLOv5l, YOLOv8s, Faster R-CNN, and SSD, respectively, and the YOLOv5s-GBR model can be used to accurately recognize overlapping or occluded apples, which can be subsequently deployed in picked robots to meet the realistic demand of real-time apple detection.

show abstract

“…Here, the local context enhancement term LCE(V) is introduced [23]. The function LCE(•) uses deep convolutional parameterization with a convolutional kernel size of 5.…”

Section: Bi-level Routing Spatial Attention Modulementioning

confidence: 99%

Detection of Orchard Apples Using Improved YOLOv5s-GBR Model

Sun,

Zheng,

et al. 2024

Agronomy

View full text Add to dashboard Cite

show abstract

“…These matrices are then combined with the LCE(.) function [14] and applied to attention. Notably, the LCE(.)…”

Section: Figure 3 Structure Of C3 C2f and Fast_c2fmentioning

confidence: 99%

Citrus Diseases and Pests Detection Model Based on Self-Attention YOLOV8

Luo,

Xue,

Deng

et al. 2023

IEEE Access

View full text Add to dashboard Cite

To accurately detect citrus diseases and pests in real time, even in complex natural environments, this study proposes a Light-SA YOLOV8 (Lightweight Self-Attention YOLOV8) model. Based on YOLOV8, the model introduces the BRA self-attention mechanism module before the SPPF layer in the backbone to overcome challenges posed by complex backgrounds, such as uneven lighting and reflections on citrus leaves and fruits, and achieve flexible computation allocation and content awareness. And, the FasterNet Block is also introduced into the backbone to decrease the computational complexity. Moreover, to enhance precision and computational efficiency for citrus diseases and pests, a new feature fusion technique, known as the AFPN (asymptotic characteristic pyramid network) structure, is employed at the Neck. The constructed dataset includes five types of diseases: anthracnose, citrus canker, melanosis, scab, and bacterial brown spot-along with one type of insect pest, the citrus shallow leaf moth. Experimental results demonstrate that the Light-SA YOLOV8 model achieves an average detection accuracy of 92.6% for the six types of diseases and pests on test set. The

show abstract

“…Use the Softmax function to complete the attention calculation on the aggregated key-value pairs. In addition, the local context enhancement (LCE) [60] function is also introduced to improve the context representation of each value with depth-wise separable convolution, of which convolution kernel size is set to 5. And the output result is expressed as follows:…”

Section: Brgatt-based Object Detector Networkmentioning

confidence: 99%

GDMR-Net: A Novel Graphic Detection Neural Network via Multi-Crossed Attention and Rotation Annotation for Agronomic Applications in Supply Cyber Security

Xue,

Zhao,

Xiu

et al. 2023

Agronomy

View full text Add to dashboard Cite

The development of smart agriculture holds great significance in ensuring the supply and cyber security of agricultural production. With the advancement of intelligent technologies, unmanned robots collaborating with the Internet of Things (IoT) play increasingly crucial roles in the realm of smart agriculture; they have become effective means to ensure agricultural safety and supply security. However, in the pursuit of unmanned agronomic applications, there is an urgent challenge: these intelligent systems generally show low accuracy in target detection when relying on visual perception due to fine-grained changes and differing postures of crops. To solve this issue, we proposed a novel multi-target detection approach via incorporating graph representation learning and multi-crossed attention techniques. The proposed model first utilizes a lightweight backbone network to accurately identify the characteristics and conditions of crops. Then, the higher-order graphic feature extractor is designed to comprehensively observe fine-grained features and potential graphic relationships among massive crops, enabling better perception capabilities of agricultural robots, allowing them to adapt to complex environments. Additionally, we can address bilevel routing by combining ghost attention and rotation annotations to handle continuous posture changes during crop growth and mutual occlusion. An extensive set of experiments demonstrated that our proposed approach outperforms various advanced methods of crop detection, achieving identification accuracies up to 89.6% (mAP) and 94.7% (AP50). Ablation studies further proved the preferable stability, of which the parameter size is only 628 Mbyte, while maintaining a high processing speed of 89 frames per second. This provides strong support for application of the technique in smart agriculture production and supply cyber security.

show abstract

Shunted Self-Attention via Multi-Scale Token Aggregation

Cited by 6 publications

References 29 publications

Detection of Orchard Apples Using Improved YOLOv5s-GBR Model

Detection of Orchard Apples Using Improved YOLOv5s-GBR Model

Citrus Diseases and Pests Detection Model Based on Self-Attention YOLOV8

GDMR-Net: A Novel Graphic Detection Neural Network via Multi-Crossed Attention and Rotation Annotation for Agronomic Applications in Supply Cyber Security

Contact Info

Product

Resources

About