基于改进YOLO v4的自然环境苹果轻量级检测方法

针对苹果采摘机器人识别算法包含复杂的网络结构和庞大的参数体量，严重限制检测模型的响应速度问题，本文基于嵌入式平台，以YOLO v4作为基础框架提出一种轻量化苹果实时检测方法(YOLO v4-CA)。该方法使用MobileNet v3作为特征提取网络，并在特征融合网络中引入深度可分离卷积，降低网络计算复杂度；同时，为弥补模型简化带来的精度损失，在网络关键位置引入坐标注意力机制，强化目标关注以提高密集目标检测以及抗背景干扰能力。在此基础上，针对苹果数据集样本量小的问题，提出一种跨域迁移与域内迁移相结合的学习策略，提高模型泛化能力。试验结果表明，改进后模型的平均检测精度为92.23%，在嵌入式平台上的检测速度为15.11f/s，约为改进前模型的3倍。相较于SSD300与Faster R-CNN，平均检测精度分别提高0.91、2.02个百分点，在嵌入式平台上的检测速度分别约为SSD300和Faster R-CNN的1.75倍和12倍；相较于两种轻量级目标检测算法DY3TNet与YOLO v5s，平均检测精度分别提高7.33、7.73个百分点。因此，改进后的模型能够高效实时地对复杂果园环境中的苹果进行检测，适宜在嵌入式系统上部署，可以为苹果采摘机器人的识别系统提供解决思路。

Lightweight Real-time Apple Detection Method Based on Improved YOLO v4

Under the picking conditions in unstructured environments, such as overlapping and occlusion, the recognition system based on deep learning in apple picking robot contained complex network structure and large parameter volumes, for which the response speed of detection model was severely limited. In response to this problem, based on the embedded platform, a lightweight apple real-time detection method called YOLO v4-CA, which selected YOLO v4 as the basic framework, was proposed. The proposed method used MobileNet v3 as the feature extraction network, and introduced deep separable convolution in the feature fusion network to reduce network computational complexity. In order to ensure the detection accuracy, coordinate attention was introduced in the key position of the network to strengthen target attention, which can improve the ability to detect dense targets and resist background interference. For the small apple datasets, a combination of cross-domain and in-domain transfer learning strategy was proposed to improve the generalization ability of the model. Experimental results showed that the average precision of the improved model was 92.23%, and the detection speed on the embedded hardware platform was 15.11 frames per second, which was about three times than that of the original YOLO v4 model. Compared with the two representative target detection algorithms of SSD300 and Faster R-CNN, the average precision was increased by 0.91 percentage points and 2.02 percentage points respectively, and the detection speed on the embedded hardware platform was about 1.75 times and 12 times that of the two respectively. Compared with the two lightweight target detection algorithms of DY3TNet and YOLO v5s, the average precision was increased by 7.33 percentage points and 7.73 percentage points respectively. Therefore, the improved model YOLO v4-CA can efficiently detect apples in a complex orchard environment in real time, and it was suitable for deployment on embedded systems. It can provide solutions for the recognition system of apple picking robots.