基于轻量化卷积神经网络的苹果目标检测算法

Lightweight Apple Object Detection Algorithm Based on Lightweight Convolutional Neural Network

针对当前苹果检测算法模型参数多、检测实时性差的问题,提出一种基于轻量化卷积神经网络的苹果目标检测算法.首先,用经典轻量化卷积神经网络ShuffleNet V2替换YOLO v5s的主干网络,实现模型的轻量化;然后,用stem模块取代主干网络的图像处理层进行初始特征提取并且嵌入SPPF结构,弥补轻量化带来的精度损失;最后,在边界框的回归损失函数中引入α幂化指标,进一步提高边界框的定位精度.试验结果表明,改进后算法模型的平均精度均值达到95.8%,网络参数量降低了85.6%,在GPU上的单张平均检测时间仅10 ms,满足苹果采摘任务对检测精度和实时性的要求.

A lightweight convolutional neural network based apple object detection algorithm is proposed to address the issues of multiple model parameters and poor real-time detection performance in current apple detection algorithms.Firstly,replace the backbone network of YOLO v5s with the classic lightweight convolutional neural network ShuffleNet V2 to achieve lightweight of the model.Secondly,in order to compensate for the accuracy loss caused by lightweight,the stem module is used to replace the image processing layer of the backbone network for initial feature extraction and embedded into the SPPF structure.Finally,anα-powered indicator is introduced into the regression loss function of the bounding box to further improve the positioning accuracy of the bounding box.The experiment shows that the average accuracy of the improved algorithm model reaches 95.8%,the number of network parameters is reduced by 85.6%,and the average detection time of a single image on GPU is only 10ms,meeting the requirements of apple picking tasks for detection accuracy and real-time performance.