基于改进YOLOv7-tiny的实时抓取检测算法

Real⁃time grasping detection algorithm based on improved YOLOv7-tiny

针对目前抓取检测任务中存在的预测位置和角度不准确、小物体适应性和检测实时性较差的问题,提出一种实时抓取检测算法。该算法以YOLOv7-tiny为基本框架,在网络特征融合阶段嵌入CA注意力机制,增强对小物体的适应性,并添加深度可分离卷积块进一步提升检测速度;引入KLD损失函数思想改进预测框回归损失函数,提高预测框位置和角度的准确性。该方法在Cornell数据集上的图像分割准确率为98.2%,对象分割准确率为97.3%,抓取检测速度达到62.5 FPS,相较其他经典方法有明显的优势,能够满足实时高精度抓取检测任务的需求。

Aiming at the problems of inaccurate position and angle prediction,adaptability of small objects and poor real-time detection in current grasping detection tasks,a real-time grasping detection algorithm is proposed.Taking YOLOv7-tiny as the basic framework,the CA attention mechanism is embedded in the network feature fusion stage to enhance the adaptability to small objects,and depth separable convolution blocks are added to further improve the detection speed.The idea of KLD loss function is introduced to improve the regression loss function of prediction frame and improve the accuracy of the position and angle of prediction frame.The image segmentation accuracy and object segmentation accuracy of this method on Cornell dataset are 98.2%and 97.3%,and the capture detection speed reaches 62.5 FPS.Compared with other classical methods,this method has obvious advantages and can meet the requirements of real-time and high-precision capture detection tasks.