关键词: ＹＯＬＯｖ４, ＫＮＮ聚类, Ｇｈｏｓｔｎｅｔ, 空间金字塔池化

Abstract:

Aiming at the problems of insufficient computing power, limited target detection speed and difficulty in meeting real-time application of edge equipment of a picking robot, this paper proposes a lightweight algorithm based on improved YOLOv4, which improves detection speed and reduces network volume, and can be better applied to edge equipment. In this paper, the lightweight backbone network Ghostnet is used to replace the CSPdarknet53 backbone network in YOLOv4. Compared with YOLOv4, it has fewer parameters and is lighter in weight, which is proposed by Huawei. When the features with common convolution are being extracted, some of the same features are merged with the network after convolution, thus reducing the amount of computation without reducing the accuracy.

On the basis of replacing the backbone network, the deep separable convolution is used to replace the convolutional block of the neck network in YOLOv4. The deep separable convolution is further optimized by being divided into two simple steps and reducing the weight and computation amount. Although the previous modified network has improved the detection speed, the accuracy has decreased. In order to improve the accuracy without greatly increasing the calculation amount and weight, the number of layers of CBL convolutional module increases before and after the space pyramid pooling, and all the three layers are replaced with five layers to increase the ability of information extraction.

In addition, better information extraction of the feature map at the end of the backbone network can improve the accuracy, so the feature extraction of the image and the information acquisition of the whole network in the image are required with the aim to improve the accuracy. In order to further improve the accuracy, KNN clustering algorithm is used to calculate the prior box for prediction so as to make better preparation for the subsequent training. The more similar the prior box is to the target box, the more accurate the network will be after training. Meanwhile, Mosaic data are used to enhance the recognition accuracy. The detection results of apple show that the modified network has a better recognition accuracy: compared with YOLOv4, the detection speed increases by 45.8%, the FPS reaches 35 frames, and the weight of the whole network reduces by 79.7%. Compared with the other algorithms introduced in this paper, the modified network is better in both accuracy and prediction speed. In comparison with the effect of the picture, the accuracy of the modified network frame selection is more accurate than that of the other networks, and the missed targets are also fewer than those of the other networks. In view of the above performance, the overall performance of the modified network is better than that of the original YOLOv4 and the other networks. The modified network improves the detection speed and reduces the weight file size, and can be better applied to the edge equipment such as picking robots with insufficient computing power and small storage space.