基于改进的YOLOv3算法在目标识别与抓取中的应用

Application of Improved YOLOv3 Algorithm in Target Recognition and Crawling

针对目标识别与抓取领域中CNN、Faster-RCNN等传统神经网络系列算法的识别准确率不高,实时性较差的问题,提出一种基于YOLOv3的改进神经网络算法。改进的YOLOv3算法主要是引用Inception网络思想,通过不同尺度的卷积核对目标进行多尺度特征提取,在增加网络宽度的同时减少YOLOv3网络的循环次数。同时,YOLOv3算法对于anchor box的选取方式使用Meanshift(均值漂移)聚类算法与K-means聚类算法相结合的方式进行改进,解决了K值需要人为测定的问题。改进的YOLOv3算法在自制数据集进行对比实验,实验结果表明:改进YOLOv3算法的mAP(Mean Average Precision)值要高于YOLOv3算法10%,在识别速度上提高了9%,在充分满足实时识别的同时提高了对中小目标识别的准确率。

The recognition accuracy of traditional neural network series algorithms such as CNN and Faster-RCNN in target recognition and capture field is not high,and the real-time performance is poor. An improved neural network algorithm based on YOLOv3 is proposed. The Inception network idea is mainly quoted in the improved YOLOv3 algorithm,which Multi-scale feature of targets is extracted by convolution kernels of different scales,the number of loops in the YOLOv3 network is reduced while increasing the network width. At the same time,the selection method of the anchor box in the YOLOv3 algorithm is improved by combining the Meanshift(mean shift) clustering algorithm and the K-means clustering algorithm;the problem that K value needs artificial measurement is solved;YOLOv3 algorithm for comparison experiments in homemade data sets is improved. Experimental results show that the m AP(Mean Average Precision) value of the improved YOLOv3 algorithm is 10% higher than the YOLOv3 algorithm;and the recognition speed is increased by 9%,which improves the accuracy of small and medium target recognition while fully satisfying the real-time recognition.