{"title":"AOD-Net:面向农业自动化的轻量级水果实时检测算法","authors":"Juntao Tong","doi":"10.1007/s11694-025-03149-1","DOIUrl":null,"url":null,"abstract":"<div><p>Fruit defect classification and quality visual inspection are crucial for automated harvesting in agriculture. To address the issues of large model parameters, low target recognition accuracy, and interference from background noise in existing detection models, we proposed a novel fruit detection algorithm, AOD-Net, which integrated Polarized self-attention (PSA) mechanism and a new lightweight structure, Cross-Stage Partial Convolution (CSPC). By adding the PSA module at the end of the backbone network, AOD-Net establishes mutual dependencies between feature channels, reducing background noise and improving the network’s ability to extract and recognize subtle target features, thus enhancing target localization accuracy. The CSPC structure, inspired by Dual-Conv and Partial-Conv, replaces certain convolutional layers, significantly reducing model parameters and accelerating detection speed while maintaining accuracy to meet real-time requirements. The Receptive-Field Attention Convolution module is incorporated into the neck network to enhance feature learning, improve feature extraction accuracy, and address parameter sharing issues, thus improving model generalization. Additionally, the Dysample upsampling operator replaces the traditional nearest-neighbor interpolation to reduce computational parameters while improving feature fusion for different fruit types, thereby enhancing detection accuracy and robustness. Experimental results on the publicly available FruitNet dataset showed that AOD-Net achieved a mAP of 93.55%, with improvements of 1.30%, 1.96%, and 3.95% in Precision, Recall, and mAP, respectively, compared to the standard YOLOv5s. The model’s memory usage decreased by 8.97%, and the computational cost was reduced from 16.0 GFLOPs to 11.3 GFLOPs, verifying the effectiveness of the proposed algorithm. AOD-Net strikes an excellent balance between speed and accuracy, making it an efficient and practical fruit detection method.</p></div>","PeriodicalId":631,"journal":{"name":"Journal of Food Measurement and Characterization","volume":"19 4","pages":"2818 - 2830"},"PeriodicalIF":2.9000,"publicationDate":"2025-02-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"AOD-Net: a lightweight real-time fruit detection algorithm for agricultural automation\",\"authors\":\"Juntao Tong\",\"doi\":\"10.1007/s11694-025-03149-1\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><p>Fruit defect classification and quality visual inspection are crucial for automated harvesting in agriculture. To address the issues of large model parameters, low target recognition accuracy, and interference from background noise in existing detection models, we proposed a novel fruit detection algorithm, AOD-Net, which integrated Polarized self-attention (PSA) mechanism and a new lightweight structure, Cross-Stage Partial Convolution (CSPC). By adding the PSA module at the end of the backbone network, AOD-Net establishes mutual dependencies between feature channels, reducing background noise and improving the network’s ability to extract and recognize subtle target features, thus enhancing target localization accuracy. The CSPC structure, inspired by Dual-Conv and Partial-Conv, replaces certain convolutional layers, significantly reducing model parameters and accelerating detection speed while maintaining accuracy to meet real-time requirements. The Receptive-Field Attention Convolution module is incorporated into the neck network to enhance feature learning, improve feature extraction accuracy, and address parameter sharing issues, thus improving model generalization. Additionally, the Dysample upsampling operator replaces the traditional nearest-neighbor interpolation to reduce computational parameters while improving feature fusion for different fruit types, thereby enhancing detection accuracy and robustness. Experimental results on the publicly available FruitNet dataset showed that AOD-Net achieved a mAP of 93.55%, with improvements of 1.30%, 1.96%, and 3.95% in Precision, Recall, and mAP, respectively, compared to the standard YOLOv5s. The model’s memory usage decreased by 8.97%, and the computational cost was reduced from 16.0 GFLOPs to 11.3 GFLOPs, verifying the effectiveness of the proposed algorithm. AOD-Net strikes an excellent balance between speed and accuracy, making it an efficient and practical fruit detection method.</p></div>\",\"PeriodicalId\":631,\"journal\":{\"name\":\"Journal of Food Measurement and Characterization\",\"volume\":\"19 4\",\"pages\":\"2818 - 2830\"},\"PeriodicalIF\":2.9000,\"publicationDate\":\"2025-02-27\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Journal of Food Measurement and Characterization\",\"FirstCategoryId\":\"97\",\"ListUrlMain\":\"https://link.springer.com/article/10.1007/s11694-025-03149-1\",\"RegionNum\":3,\"RegionCategory\":\"农林科学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q2\",\"JCRName\":\"FOOD SCIENCE & TECHNOLOGY\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Journal of Food Measurement and Characterization","FirstCategoryId":"97","ListUrlMain":"https://link.springer.com/article/10.1007/s11694-025-03149-1","RegionNum":3,"RegionCategory":"农林科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"FOOD SCIENCE & TECHNOLOGY","Score":null,"Total":0}
AOD-Net: a lightweight real-time fruit detection algorithm for agricultural automation
Fruit defect classification and quality visual inspection are crucial for automated harvesting in agriculture. To address the issues of large model parameters, low target recognition accuracy, and interference from background noise in existing detection models, we proposed a novel fruit detection algorithm, AOD-Net, which integrated Polarized self-attention (PSA) mechanism and a new lightweight structure, Cross-Stage Partial Convolution (CSPC). By adding the PSA module at the end of the backbone network, AOD-Net establishes mutual dependencies between feature channels, reducing background noise and improving the network’s ability to extract and recognize subtle target features, thus enhancing target localization accuracy. The CSPC structure, inspired by Dual-Conv and Partial-Conv, replaces certain convolutional layers, significantly reducing model parameters and accelerating detection speed while maintaining accuracy to meet real-time requirements. The Receptive-Field Attention Convolution module is incorporated into the neck network to enhance feature learning, improve feature extraction accuracy, and address parameter sharing issues, thus improving model generalization. Additionally, the Dysample upsampling operator replaces the traditional nearest-neighbor interpolation to reduce computational parameters while improving feature fusion for different fruit types, thereby enhancing detection accuracy and robustness. Experimental results on the publicly available FruitNet dataset showed that AOD-Net achieved a mAP of 93.55%, with improvements of 1.30%, 1.96%, and 3.95% in Precision, Recall, and mAP, respectively, compared to the standard YOLOv5s. The model’s memory usage decreased by 8.97%, and the computational cost was reduced from 16.0 GFLOPs to 11.3 GFLOPs, verifying the effectiveness of the proposed algorithm. AOD-Net strikes an excellent balance between speed and accuracy, making it an efficient and practical fruit detection method.
期刊介绍:
This interdisciplinary journal publishes new measurement results, characteristic properties, differentiating patterns, measurement methods and procedures for such purposes as food process innovation, product development, quality control, and safety assurance.
The journal encompasses all topics related to food property measurement and characterization, including all types of measured properties of food and food materials, features and patterns, measurement principles and techniques, development and evaluation of technologies, novel uses and applications, and industrial implementation of systems and procedures.