基于改进YOLOv4的轻量化菠萝苗心检测算法

当前菠萝催花作业以人工喷洒为主,生产效率低、劳动强度大。菠萝苗心位置的精准识别和定位是实现机械化、智能化菠萝催花的核心问题。该研究在YOLOv4目标识别算法的基础上,选择GhostNet作为主干特征提取网络,构建了一种混合网络模型,并在颈部网络中融合深度可分离卷积与轻量级的注意力模块。改进后的模型相较于YOLOv4模型的总参数量减少70%。与YOLOv4、Faster R-CNN和CenterNet 3个模型进行检测对比试验,结果可得：改进模型在菠萝植株种植密集与稀疏的条件下识别精度分别为94.7%和95.5%,实时识别速度可达27帧/s,每张图像平均检测时间为72 ms,相比常规YOLOv4模型用时缩短23%。总体性能表现均优于对比组的目标检测模型。总的来说,改进模型YOLOv4-GHDW在一定程度上实现了检测速度、识别精度和模型体量三者之间平衡,能够在实际种植环境下对菠萝苗心有较好的识别效果。研究结果可为智能化菠萝精准催花设备研发提供视觉技术支持。

Lightweight algorithm for pineapple plant center detection based on improved an YOLOv4 model

Abstract: Spraying is one of the main physiological and biochemical processes for flower induction in pineapple production. However, manual spraying cannot fully meet large-scale production in recent years, due to the very high labor intensity. Fortunately, automatic spraying can be expected for the flower induction of pineapple during this time. Among them, it is a high demand to accurately detect the center position of pineapple plants during spraying. In this study, a hybrid network model was proposed to accurate recognize the pineapple plant center using an improved YoloV4 deep learning. The conventional YoloV4 model shared an outstanding detect speed more suitable for the agricultural robot. However, the large number of parameters occupied a huge amount of memory in the original model, leading to more computing power in the device. Therefore, it was necessary for a much lighter and more flexible network structure in the reasoning work, particularly with the shorter running time in the actual field operation. The lightweight was then realized using the network framework of the YoloV4 model. Firstly, the number of network layers and feature map channels in the backbone were reduced using the GhostNet network structure as the backbone to replace the CSP-Darknet53, which enhanced the efficiency of the feature extraction. The GhostNet presented a Ghost module for building efficient neural architectures, in order to reduce the computational costs of recent deep neural networks. As such, the Ghost bottlenecks performed better in the lightweight structures of the entire network. Secondly, the depthwise separable convolution was introduced into the Neck network. The convolution set was integrated to reduce the number of parameters with a high level of feature extraction. Besides, the lightweight attention mechanism (called CBAM) was added between the backbone and the neck. Attention weights were then put on the channel and spatial dimensions for a better connection in the channel and space, in order to extract the effective features of the target. Finally, the quantity of the improved model (named Yolov4-GHDW) parameters was reduced by 70%, compared with the original. In the dataset-making stage, some operations were selected to process the collected images for the better generalization ability of the training model, including the vertical flip, adding Gaussian blur, adding salt and pepper noise, and extracting green features. Among them, 2530 images were divided into 1962, 218, and 350 images for the training set, the validation set, and the test set. The collected image dataset was used for the model training at the workstation, where the loss value of the training set and test set were recorded. The model with the smallest loss value was then selected as the optimal solution after training. The results show that the Yolov4-GHDW model presented a much higher recognition speed, and an excellent extraction performance to the target, compared with the original YoloV4, Faster R-CNN, and CenterNet. Specifically, the average accuracies of target recognition under the dense and sparse planting of pineapple plants were 94.7% and 95.5%, respectively, with only 68.4 MB of model memory. The real-time recognition speed reached 27 frames/s, and the average inference time of each image was 72ms, which was 23% shorter than that of YoloV4. Over all, the improved model can realize the identification and detection of pineapple plant center under complex conditions. The performance can fully meet the location requirement of the equipment during operation. Moreover, the improved model was also deployed in an embedded device (named Jetson Nano) to verify the performance. As a result, 13 frames per second was achieved to fluently operate with low power consumption. In conclusion, the Yolov4-GHDW has a high accuracy and recognition speed with a small memory footprint, indicating the balance between speed and accuracy. The finding can also provide technical support for intelligent and accurate flowering equipment in pineapple production.