基于改进YOLOv7的遮挡番茄定位识别算法

Occluded Tomato Localization Recognition Algorithm based on improved YOLOv7

关键词:  YOLOv7  注意力机制  小目标检测  遮挡番茄

Key words:  YOLOv7  Attention mechanism  Small target detection  Occluded tomato

摘要:

为提高番茄采摘机器人对遮挡番茄果实的识别准确率，提出了一种改进YOLOv7的遮挡番茄定位识别算法OT-YOLO（YOLOv7 for Occluded Tomatoes）。为了在保持模型轻量化的同时增强对遮挡目标番茄果实的感知能力，通过融入深度可分离卷积和无参数的注意力机制（SimAM）改进原特征提取网络中的MP模块；其次，在颈部网络和头部网络分别引入SE通道注意力机制，以减少无效目标的干扰，加强对遮挡目标的关注，提升模型的检测精度；为了进一步增强对遮挡小番茄果实的识别提取能力，在头部网络增加极小目标检测层。实验结果表明，OT-YOLO模型在Tomato数据集上的平均检测精度、精确率和召回率分别达到93.8%、98%和98%，相比YOLOv7算法，分别提高了2.7%、2.4%和4%，模型参数量降低了2.5 M，拥有更小的模型参数与更好的识别效果。

Abstract:

In order to improve the recognition accuracy of tomato picking robots for occluded tomato fruits, an improved YOLOv7 for Occluded Tomatoes Localization and Recognition Algorithm OT-YOLO (YOLOv7 for Occluded Tomatoes) is proposed. In order to enhance the ability to perceive the occluded target tomato fruit while reducing the module parameters to keep the model lightweight, the MP module in the original feature extraction network is improved by incorporating depth-separable convolution and parameter-free attentional mechanism (SimAM). Second, the SE channel attention mechanism is introduced into the neck network and the head network, respectively, in order to reduce the interference of invalid targets, enhance the attention to the occluded targets, and improve the detection accuracy of the model. In order to further enhance the ability of recognizing and extracting occluded small tomato fruits, a tiny target detection layer is added to the head network. The experimental results show that the average accuracy mAP@.5, precision and recall of OT-YOLO model on Tomato dataset reaches 93.8%, 98% and 98%, respectively, which are 2.7%, 2.4% and 4% higher compared to the YOLOv7 algorithm, and the number of model parameters reduces by 2.5M, possessing smaller model parameters with better recognition results.