基于改进YOLO及NMS的水果目标检测

Fruit target detection based on improved YOLO and NMS

为使水果采摘机器人在复杂情况下如树叶遮挡、果实目标尺度变化大等情况能准确地检测出水果,提出一种YOLO(you only look once)改进模型与NMS(non-maximum suppression)改进算法的目标检测方法。首先,对传统YOLO深度卷积神经网络架构进行改进,设计一种更细化的SPP5(spatial pyramid pooling)特征融合网络模块,强化特征图多重感受野信息的融合,并基于此模块提出一种YOLOv4-SPP2-5模型,在标准YOLOv4网络中跨层添加并改进SPP层,重新分布池化核大小,增强感受野范围,从而降低目标误检率;其次,提出一种Greedy-Confluence的NMS改进算法,通过对高度接近的检测框直接抑制和对重叠检测框综合考虑距离交并比DIOU(distance-intersection over union)和加权接近度WP(weighted proximity)的方法,均衡NMS的计算消耗并减少检测框的错误抑制,从而提高遮挡、重叠物体的检测精度;最后,分别对改进方法进行性能测试,验证方法的可行性,随后制作水果检测数据集并进行格式转换和标签标注,然后采用数据增强技术对训练数据进行扩充,并使用K-means++聚类方法获取先验锚定框,在计算机上进行了水果检测实验。结果表明,基于改进YOLO网络及改进NMS的水果检测方法在准确率方面有显著的提高,平均精度均值(mean average precision,MAP)在YOLOv4上达到了96.65%,较原网络提升1.70%,并且实时性也得到了保证,在测试设备上达到了39.26帧/s。

To enable fruit picking robots that accurately detect target in complex conditions such as leaf covering and variances of fruit sizes,etc.,improved YOLO(you only look once)model and NMS(non-maximum suppression)algorithm are proposed.First,the traditional YOLO deep convolutional neural network architecture is upgraded.A more fine-grained SPP5(spatial pyramid pooling)feature fusion network module is generated to enhance the integration of multiple sensory field information in feature maps,based on which a YOLOv4-SPP2-5 model is proposed.The SPP layer is added and improved in the standard YOLOv4 network across layers and redistributed the pooling kernel size and to enlarge perceptual field range,thus decreasing the false detection rate.Moreover,an improved Greedy-Confluence NMS algorithm is proposed.Through direct suppression of high-proximity detection boxes and comprehensive consideration of Distance-Intersection over Union(DIOU)and weighted proximity(WP)for overlapping detection boxes,the computational consumption of NMS was balanced and the error suppression of detection boxes was reduced,so as to improve the detection accuracy of occlusion and overlapping objects.Finally,performance tests are conducted to verify the feasibility of the method,followed by format converting and annotation labelling with fruit training datasets.The training datasets are expanded via data augmentation techniques and the K-means++clustering approach is utilized to obtain a priori anchor frames,and the fruit detection experiments are carried out on a computer.The results demonstrate that the improved YOLO network and NMS algorithm-based approach significantly increase the accuracy rate of fruit detection.The mean average precision(MAP)reaches 96.65%at YOLOv4,which is 1.70%higher than the previous network.Real-time performance is also guaranteed,hitting 39.26 frames per second on the test device.