Pomelo Tree Detection Method Based on Attention Mechanism and Cross-Layer Feature Fusion

Abstract

Deep learning is the subject of increasing research for fruit tree detection. Previously developed deep-learning-based models are either too large to perform real-time tasks or too small to extract good enough features. Moreover, there has been scarce research on the detection of pomelo trees. This paper proposes a pomelo tree-detection method that introduces the attention mechanism and a Ghost module into the lightweight model network, as well as a feature-fusion module to improve the feature-extraction ability and reduce computation. The proposed method was experimentally validated and showed better detection performance and fewer parameters than some state-of-the-art target-detection algorithms. The results indicate that our method is more suitable for pomelo tree detection.

Detecting and Mapping Individual Fruit Trees in Complex Natural Environments via UAV Remote Sensing and Optimized YOLOv5

“…As mentioned in Section 1, in our previous study, we proposed a YOLOx-nano pomelo tree detection method based on an attention mechanism and cross-layer feature fusion and showed that this method was more suitable for pomelo tree detection than was other state-of-the-art object detection algorithms. The present study showed that YOLOv5s and its attention-optimized models can detect IPTs with high accuracy, in line with the results of Yuan et al [49]. Although the structure of the network proposed by Yuan et al [49] was lightweight, the AP value reached 93.74%.…”

“…The present study showed that YOLOv5s and its attention-optimized models can detect IPTs with high accuracy, in line with the results of Yuan et al [49]. Although the structure of the network proposed by Yuan et al [49] was lightweight, the AP value reached 93.74%. Our optimized YOLOv5s models fully outperformed this network in terms of AP, which were all higher than 94.00%, with the highest AP value of 94.50%.…”

“…However, most of those previous works needed to improve the architecture of CNN networks to increase the performance of deep learning models. Yuan et al [49] proposed an improved YOLOx-nano algorithm to detect pomelo trees and compared it with several state-of-the-art object detection algorithms, such as Faster R-CNN, SSD, YOLOv3, and YOLOv4-tiny. Their method showed better suitability for pomelo tree detection in UAV images with better performance and fewer parameters.…”

Optimized anchor-free network for dense rotating object detection in remote sensing images

“…In the training process of large-scale remote sensing image object extraction algorithms, CNNs are usually used. Due to the limited size of shared convolutional kernels involved in network operations and not changing based on the size of the extracted object in the task, its global modeling ability is limited, which weakens the connection between the object to be extracted and its background in the image, resulting in the loss of some implicit spatial relationship information 52 , 53 . In the process of purifying and fusing feature information using feature extraction networks, the number of pixels per unit area decreases exponentially with the size of the feature map, resulting in an increase in the object information represented by a single pixel.…”

“…Due to the limited size of shared convolutional kernels involved in network operations and not changing based on the size of the extracted object in the task, its global modeling ability is limited, which weakens the connection between the object to be extracted and its background in the image, resulting in the loss of some implicit spatial relationship information. 52,53 In the process of purifying and fusing feature information using feature extraction networks, the number of pixels per unit area decreases exponentially with the size of the feature map, resulting in an increase in the object information represented by a single pixel. When the objects are densely arranged, due to the continuous feature information abstraction, interference and mosaic phenomena will appear between the feature information of adjacent objects, resulting in dense object features being difficult to distinguish and interfering with each other.…”