| The fruit accumulates nutrients such as sugar during the young fruit stage,with significant changes in surface area,transverse and longitudinal diameter and volume.The young fruits accurate detection technology provides an effective means of long-term,high-frequency and non-destructive recording data on fruit growth indicators,which helps to improve the level of information management and control in orchards and reduce human and material input.The automatic acquisition of young fruit phenotype data also provides basic data for fruit tree genetic breeding research,which has important research significance and economic benefits.In order to solve the detecting problem of complex environmental factors,near scenery color and different scale objects in unstructured environments,young apples grown at the test site affiliated with the College of Horticulture in Northwest A&F University were taken as the research object in this study,and the research on fruit detection methods based on Convolutional neural network（CNN）and visual attention mechanism was carried out,so as to lay a foundation for establishing young fruit growth prediction models.The findings of the study are as follows:（1）For the problem of low-quality imaging caused by natural factors such as wind and light in unstructured open environment,the low-quality young fruit detection method was proposed,which combined YOLOv4 model with the Non-local attention module（NLAM）and the Convolutional block attention module（CBAM）.The experimental results showed that the detection precision,recall rate,average precision and F1score of the proposed YOLOv4-NLAM-CBAM model were 85.8%,97.3%,97.2%and 91.2%,respectively,and the average run time was 35.1 ms.For highlight/shadow,blur,severe occlusion and other images,the average precision of the proposed algorithm was 98.0%,96.2%,97.0%and 96.9%,respectively.Compared with YOLOv3,the average precision of the proposed algorithm for the four types of images was improved by 1.9%,2.9%,4.5%and 3.7%,respectively.Compared with YOLOv3-SPP,the average precision of the proposed algorithm for the four types of images was decreased by 0.2%,increased by 1.2%,increased by 0.8%and increased by 1.6%,respectively.Compared with YOLOv4,the average precision of the proposed algorithm for the four types of images was increased by 0.3%,increased by 0.4%,equal and decreased by 0.1%,respectively.It showed that the proposed method in this paper can effectively improve the ability of the model to extract young apple features in low-quality images and provide some reference for intelligent monitoring of fruit growth state and growth prediction.（2）For the problem that young apples are highly similar in colour to the background of branches and leaves and easily confused with the background environment causing difficulty in detection,the deep learning method was proposed that fuses the Squeeze-and-excitation block（SE block）and Non-local block（NL block）in the architecture of deep CNN.It avoid irrelevant background information interference by enhancing the channel and non-local features.The test set experiment results showed that the average precision of the method was96.9%,which was 6.9%,1.5%and 0.2%higher than that of SSD,Faster R-CNN and YOLOv4 model,respectively.The results of the three ablation experiments on validation set showed that the SE block and NL block led to different degrees of precision improvement,the structure was well designed,and enhanced the young apple feature perception ability of the model with a small increase in the number of parameters.It showed that the proposed method is effective and feasible to achieve near-scenery young fruit detection and can provide reference for green fruit detection research.（3）For the problem that tiny fruit occupies less pixels and effective information is easily lost in deep convolution,YOLOX-S network was adopted to achieve tiny young apple detection.The experiment results showed that the average precision of the model was 91.7%,which was 6.1%and 15.6%higher than that of Center Net and YOLOv4-Tiny model,respectively.The model had 363 network layers,with 26.6 GFLOPs of floating point operations,8.9 M parameters and 101.8 f/s of frame rate.The results showed that the algorithm has excellent performance in operation efficiency and fruit detection accuracy,and it can achieve end-to-end recognition and detection of faint targets in unstructured environments.The research results are of great significance for large-scale and intelligent precision planting.（4）The software system of young apple detection in unstructured environment was developed by Py QT5 user interface compilation toolbox and Python programming language.The software system was mainly based on the low-quality young apple image detection method based on YOLOv4 network model and visual attention mechanism,the near-scenery young apple detection method based on improved deep CNN with visual attention mechanism,and the tiny young apple detection method based on YOLOX-S network,and integrated and demonstrated the young apple detection process in different scenes.The experiments showed that the software can realize the functions of young apple detection and visualization based on the above three algorithms,and the software interface has the advantages of clear design,simple and intuitive operation,clear image display and so on. |
