Shang Fengnan, Zhou Xuecheng, Liang Yingkai, Xiao Mingwei, Chen Qiao, Luo Chendi. Dragon fruit detection method in natural environment based on improved YOLOX. Smart Agriculture, 2022, 4: 120-131.

citation format

Shang Fengnan, Zhou Xuecheng, Liang Yingkai, Xiao Mingwei, Chen Qiao, Luo Chendi. pitaya detection method in natural environment based on improved YOLOX [J]. Smart Agriculture (Chinese and English), 2022, 4(3): 120-131.

SHANG Fengnan, ZHOU Xuecheng, LIANG Yingkai, XIAO Mingwei, CHEN Qiao, LUO Chendi. Detection method for dragon fruit in natural environment based on improved YOLOX[J]. Smart Agriculture, 2022, 4(3): 120-131.

CNKI reading

official website reading (full text, free)

Detection method of dragon fruit in natural environment based on improved YOLOX

Shang Fengnan ^{1, 2, 3}, Zhou Xuecheng ^{1, 2, 3*}, Liang Yingkai ^{1, 2, 3}, Xiao Mingwei ^{1, 2, 3}, Chen Qiao ^{1, 2, 3}, Luo Chendi ^{1, 2, 3}

(1. South China Agricultural University College of Engineering , Guangzhou, Guangdong 510642; 2. Guangdong Provincial Key Laboratory of Agricultural Artificial Intelligence, Guangzhou, Guangdong 510642; 3. Key Laboratory of Key Technologies of Southern Agricultural Machinery and Equipment, Ministry of Education, Guangzhou 510642)

Abstract: Accurate detection of fruits in the natural environment is a prerequisite for dragon fruit picking robots to perform picking operations. In order to improve the accuracy, robustness and detection efficiency of fruit recognition in natural environments, this study improves the YOLOX (You Only Look Once X) network and proposes a target detection method containing an attention module. In order to facilitate deployment on embedded devices, this method uses the YOLOX-Nano network as the benchmark, and adds the Convolutional Block Attention Module (CBAM) to the backbone feature extraction network of YOLOX-Nano. It learns the correlation of features between different channels by assigning weight coefficients to feature layers of different scales extracted from the backbone network, strengthens the transmission of deep information in the network, and reduces the interference to dragon fruit recognition in the natural environment background. Performance evaluation and comparative tests were conducted on this method. After training, the AP_0.5 value of the dragon fruit target detection network in the test set was 98.9%, and the AP_0.5:0.95 value was 72.4%. Comparing other YOLO network models under the same experimental conditions, the average detection accuracy of this method exceeds the YOLOv3, YOLOv4-Tiny and YOLOv5-S models by 26.2%, 9.8% and 7.9% respectively. Finally, real-time testing was conducted on videos collected in the natural environment of dragon fruit orchards with different resolutions. The test results show that the improved YOLOX-Nano target detection method proposed in this study has an average detection time of 21.72 ms per frame, a F₁ value of 0.99, and a model size of only 3.76 MB. The detection speed, detection accuracy, and model size meet the technical requirements for dragon fruit picking in the natural environment.

Keywords: fruit picking; natural environment; dragon fruit; target detection; YOLOX; attention mechanism; deep learning YOLOX

Fig. 3 Convolutional attention module structure

Fig. 3 Convolutional Block Attention Module structure

Fig. 4 Loss curves of YOLOX-Nano with different input resolutions

Fig. 5 Comparison of test results of YOLOX-Nano before and after improvement

Fig. 6 Comparison of dragon fruit detection results before and after YOLOX-Nano improvement

Fig. 6 Test results comparison of YOLOX-Nano before and after the improvement

Fig. 7 Dragon fruit detection effect of different networks during lighting

Fig. 8 Dragon fruit detection effect of different networks during shading

Fig. 9 Dragon fruit detection effect of different networks during backlighting

Fig. 10 Comparison of dragon fruit detection frame rates of improved YOLOX-Nano network

Corresponding author

Zhou Xuecheng Professor

Zhou Xuecheng, Professor of at South China Agricultural University, PhD in Engineering. He presides over the national "863" program, the National Key Research and Development Program, and the National Natural Science Foundation of China General Project. His research work involves intelligent imaging detection and intelligent agricultural machinery and equipment. His main research directions include three-dimensional imaging detection of in-situ root systems, non-destructive testing of internal quality of agricultural products, and fruit and vegetable picking robots. He has published more than 30 academic papers and obtained 34 nationally authorized invention patents and 5 software copyrights.

Source: "Smart Agriculture (Chinese and English)" Issue 3, 2022

Please contact the editorial department for authorization for reprinting

Supporting unit of this issue

Zhejiang Zhenshan Technology Co., Ltd.

Weichai Lovol Heavy Industry Co., Ltd.

Small

store

Welcome to purchase

Smart Agriculture WeChat Communication Service Group

In order to facilitate academic exchanges among readers, authors and review experts in the field of agricultural science, and promote the development of smart agriculture , and to better serve readers, authors and reviewers, the editorial department has established a WeChat communication service group. Discussions on issues in the professional field and issues related to submission can be consulted in the group. How to join the group: Add the editor on WeChat 331760296, remarks: name, unit, research direction, the editor will invite you to join the group, institutional marketing and advertising personnel please do not disturb.

Publish and collect information

Welcome to publish information about the scientific research team, innovative scientific research results and related activities on my public account.