基于改进YOLOv7的猕猴桃果实识别优化

Optimization of Actinidia chinensis Fruit Recognition based on Improved YOLOv7

为了解决猕猴桃Actinidia chinensis果实识别过程中存在果实之间重叠导致的遮挡严重、检测结果易受叶片影响等问题,建立不同日照条件下的猕猴桃果实图像数据集,对YOLOv7模型做了3方面改进:将Backbone部分的卷积模块替换成GhostConv模块,在维持原有精度的程度上降低模型的参数量;针对猕猴桃果实之间存在大量重叠的情况,引入非极大值抑制NMS(Soft-NMS)策略提高检测框回归精度;融合SimAM注意力机制,增强模型对于高密度猕猴桃特征的提取能力。通过对比实验表明,优化后的模型与Faster RCNN相比,mAP值增加了12.7个百分点,检测速度提升106.8帧/s,综合性能较好,满足机器实时对于猕猴桃果实识别的需求。

In order to solve the problems of severe occlusion caused by overlapping fruits and susceptibility to leaf influence in the recognition process of Actinidia chinensis fruit,A.chinensis fruit image dataset was established under different sunlight conditions.Three improvements were made to the YOLOv7 model:replacing the convolutional module of the Backbone part with the GhostConv module,reducing the number of model parameters while maintaining the original accuracy;to address the significant overlap between A.chinensis fruits,a Non Maximum Suppression NMS(Soft NMS)strategy is introduced to improve the accuracy of detection box regression;integrating SimAM attention mechanism to enhance the models ability to extract high-density A.chinensis fruit features.Through comparative experiments,it was shown that the optimized model increased mAP value by 12.7%and detection speed by 106.8 frames/s compared to Faster RCNN.The overall performance is good and meets the real-time recognition needs of machines for A.chinensis fruit.