果园苹果的快速准确识别是实现果实产量预估和自动化采摘的关键技术之一。然而，在自然场景下，无人机拍摄苹果果实容易出现枝叶遮挡、苹果目标小等问题，导致识别精度较低。而且，当前的目标检测模型精度高但参数量较庞大，对终端设备的算力要求高。若模型部署在性能较低的自动化采摘设备中，将会出现检测实时性差的问题。针对上述问题，本文基于深度学习技术，研究自然场景下无人机图像中苹果果实识别以及轻量化的算法。

本文的主要研究内容如下：

(1) 针对果园自然场景下无人机图像存在果实密集重叠、相互遮挡、光照不均匀的问题，导致苹果识别精度低等问题，提出了改进YOLOv8的无人机图像苹果果实识别方法。采用通用高效层聚合网络(Generalized Efficient Layer Aggregation Network，GELAN)替换了YOLOv8主干网络中的C2f模块，有效提取苹果的细节特征且减少了识别模型的参数量。将动态大卷积核空间注意机制(Large Separable Kernel Attention，LSKA )添加到YOLOv8的空间金字塔池化模块的后端，以提取更丰富的特征。在Neck层中引入上采样算子DySample增强多尺度特征融合能力。在Head层添加一个尺度为160×160的小目标检测头，增强对遮挡小目标的识别能力。在自建数据集上的实验结果表明，改进后的苹果识别模型准确率为92.0%，召回率为91.2%，平均精度均值为96.6%。与原YOLOv8模型相比分别提升了1.8个百分点，2.5个百分点和1.6个百分点。该方法对于自然场景下无人机图像重叠小目标苹果具有准确的识别效果。

(2) 针对改进后的苹果果实识别模型参数多且计算量大，不适用于移动端部署等问题，对LP-YOLOv8苹果识别模型进行轻量化设计。采用C3Ghost卷积轻量化Neck层的C2f模块，降低模型参数量和计算量同时，提高识别速度。为了减轻轻量化后对苹果识别模型精度的影响，采用部分卷积(Partial Convolution，PConv)替换C2f模块中的BottleNeck结构，引入LP-YOLOv8模型主干网络替换GELAN。同时，在轻量化模型主干网络后端添加自注意力Non-Local以提高模型对苹果的特征提取能力，增加检测精度。实验结果表明，轻量化后的苹果识别模型准确率、召回率和平均精度分别为91.1%、91.6%、96.5%。相比原YOLOv8模型内存占用量减小19.4%，准确率、召回率和平均精度分别提升0.9、2.9和1.5个百分点。与当前经典的轻量化目标检测算法相比，模型性能优于其它轻量化方法，达到了轻量化的识别要求。

本文提出的无人机图像苹果果实识别模型能提升识别性能，且模型体积小，能实时识别，具有实际的应用价值。研究成果可为果实信息的自动获取提供一种有效手段，也为实现果园生产管理的自动化和智能化奠定基础。

Fast and accurate identification of orchard apples is one of the key technologies to achieve fruit yield prediction and automated picking. However, in natural scenarios, UAV photography of apple fruits is prone to problems such as branch and leaf obstruction and small apple targets, resulting in low recognition accuracy. Moreover, current target detection models have high accuracy but large number of parameters, which requires high arithmetic power of terminal devices. If the model is deployed in automated picking equipment with low performance, the problem of poor detection real-time will occur. Aiming at the above problems, this thesis is based on deep learning technology to study the algorithms for apple fruit recognition as well as lightweighting in UAV images in natural scenes.

The main research content of this thesis is as follows:

(1) Aiming at the problems of dense and overlapping fruits, mutual occlusion, and uneven illumination in UAV images in natural scenes of orchards, which lead to low accuracy of apple recognition, an improved apple fruit recognition method for UAV images with YOLOv8 is proposed. The Generalized Efficient Layer Aggregation Network (GELAN) is used to replace the C2f module in the YOLOv8 backbone network, which effectively extracts the detailed features of apples and reduces the number of parameters in the recognition model. A dynamic Large Separable Kernel Attention (LSKA) mechanism is added to the back-end of the spatial pyramid pooling module of YOLOv8 to extract richer features. The up-sampling operator DySample is introduced in the Neck layer to enhance the multi-scale feature fusion capability. Add a small target detection head with a scale of 160×160 in the Head layer to enhance the recognition of occluded small targets. Experimental results on the self-built dataset show that the improved apple recognition model has an accuracy of 92.0%, a recall of 91.2%, and an average precision mean of 96.6%. This is an improvement of 1.8 percentage points, 2.5 percentage points and 1.6 percentage points respectively compared to the original YOLOv8 model. The method has accurate recognition effect for overlapping small target apples in UAV images in natural scenes.

(2) Aiming at the problems that the improved apple fruit recognition model has many parameters and large computation volume, which is not suitable for mobile deployment, the LP-YOLOv8 apple recognition model is designed to be lightweight. The C3Ghost convolution is used to lighten the C2f module of the Neck layer, which reduces the number of model parameters and computation volume and improves the recognition speed at the same time. In order to reduce the impact of lightweighting on the accuracy of the apple recognition model, partial convolution (PConv) is used to replace the BottleNeck structure in the C2f module, and the LP-YOLOv8 model backbone network is introduced to replace the GELAN. at the same time, a self-attention Non-Local is added to the backend of the lightweight model backbone network to improve the feature extraction of the model for apples. to improve the feature extraction ability of the model for apples and increase the detection accuracy. The experimental results show that the accuracy, recall and average precision of the lightweighted apple recognition model are 91.1%, 91.6% and 96.5%, respectively. Compared with the original YOLOv8 model, the memory usage is reduced by 19.4%, and the accuracy, recall and average precision are improved by 0.9, 2.9 and 1.5 percentage points, respectively. Compared with the current classical lightweight target detection algorithms, the model performance is better than other lightweight methods and meets the lightweight recognition requirements.

The UAV image apple fruit identification model proposed in this thesis can improve the identification performance, and the model is small in size and can be identified in real time, which has practical application value. The research results can provide an effective means for automatic acquisition of fruit information, and also lay the foundation for realising the automation and intelligence of orchard production management.