An improved YOLOv8 method for identifying empty cell and unqualified plug seedling

作     者：Li, Lei Yu, Jiajia Lu, Yujun Gu, Yue Liang, Sijia Hao, Shuai 

作者机构：Zhejiang Sci Tech Univ Hangzhou 310018 Zhejiang Peoples R China Zhejiang Inst Mech & Elect Engn Hangzhou 310053 Zhejiang Peoples R China Zhejiang Modern Agr Equipment Design & Res Inst Hangzhou 310003 Zhejiang Peoples R China 

出 版 物：《JOURNAL OF REAL-TIME IMAGE PROCESSING》 (J. Real-Time Image Process.)

年 卷 期：2024年第21卷第6期

页      面：188页

核心收录：

学科分类：0808[工学-电气工程] 1002[医学-临床医学] 08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：Zhejiang Provincial Science and Technology Program [2022C02056] Zhejiang Province "Three Rural Nine Parties" Science and Technology Collaboration Program [2024SNJF069] 

主　　题：Deep learning Seedling recognition YOLOv8s Lightweighting Attention mechanism 

摘      要：A lightweight seedling detection model with improved YOLOv8s is proposed to address the seedling identification problem in the replenishment process in industrial vegetable seedling *** CBS module for feature extraction in Backbone and Neck has been replaced with a lightweight depthwise separable convolution (DSC) in order to reduce the number of model parameters and increase the speed of detection. Furthermore, the fifth layer of Backbone has been augmented with efficient multiscale attention (EMA), which can aggregate multi-scale spatial structure information more rapidly through the two branches of the parallel structure, thereby enhancing the extraction of multi-scale features. Ultimately, the computational complexity of the model is further reduced by enhancing the structure of the bottleneck to form the VoVGSCSP module, which replaces the C2f module in Neck. The mAP of the improved model on the test set is 96.2%, its parameters, GFLOPS, and model size are 7.88 M, 20.9, and 16.1 MB, respectively. The detection speed of the algorithm is 116.3 frames per second (FPS), which is higher than that of the original model (107.5 FPS). The results indicate that the improved model can accurately identify empty cell and unqualified seedling in the plug tray in real time and has a smaller number of parameters and GFLOPS, making it suitable for use on embedded or mobile devices for seedling replenishment and contributing to the realization of automated and unmanned seedling replenishment.