改进YOLOv4-tiny网络的日用商品目标检测算法

AN IMPROVED DAILY COMMODITY TARGET DETECTION ALGORITHM BASED ON YOLOV4-TINY NETWORK

针对基于移动平台的商品图像检测算法存在硬件要求高、模型复杂且精度低等问题,提出一种YOLOv4-tiny改进网络,减少网络参数与模型尺寸,提高网络精度,构建更高效的网络。将原有标准卷积替换为点卷积与逐深度卷积,特征提取使用CG模块,降低网络模型计算损耗。特征融合时,在原有特征金字塔(Feature Pyramid Networks,FPN)基础上添加PANity模块,缩短高低间卷积层的跨度。使用CSPConcat结构对此前各层融合特征进行特征优化处理,提高各层间特征融合的能力。利用k-prototypes算法优化日用商品数据集先验框的尺寸与数目。通过在darknet深度学习框架下,对日用商品数据集进行实验,得出改进后的算法平均精度(mAP)为98%,召回率为97%,较原网络提升了2.4百分点和2百分点,网络模型计算量较原网络降低了40.4%,模型存储文件缩小了55.9%。改进后的网络模型更轻量化、准确率更高,更加适用于部署在无人结算环节的低硬件水平嵌入式设备中。

For the problems of high hardware requirements,complex model and low accuracy of commodity image detection algorithm based on mobile platform,an improved network based on YOLOv4-tiny is proposed,which can reduce the network parameters and model size,improve the network accuracy and build a more efficient network.The original standard convolution was replaced by point convolution and depth convolution,and CG module was used for feature extraction to reduce the calculation loss of network model.In feature fusion,PANity module was added to the original feature pyramid(FPN)to shorten the span of convolution layer between high and low.The CSPConcat structure was used to optimize the fusion features of each layer,which improves the ability of feature fusion.k-prototypes algorithm was used to optimize the size and number of prior boxes in daily commodity data set.Through the experiment on the daily commodity data set under the framework of Darknet deep learning,it is concluded that the average accuracy(map)of the improved algorithm is 98%,the recall rate is 97%,which is 2.4 and 2 percentage points higher than the original network,the calculation amount of the network model is 40.4%lower than the original network,and the storage file of the model is 55.9%smaller.The experimental results show that the improved network model is lighter and more accurate,which is more suitable for low hardware level embedded devices deployed in unmanned settlement link.