Tomatoes Maturity Detection Approach Based on YOLOv5 and Attention Mechanisms

Abstract

Tomato is one of the most common vegetables in the world. Tomato maturity detection is important to improve tomato production. To achieve the normal operation of tomato picking robots and fast detection of tomato maturity in complex environments, a YOLOv5 feature fusion network based on attention module is proposed in this paper. By adding the attention module in the feature fusion stage, the feature information of different scales is fused, which improves the detection accuracy and reduces the weight of non-important features. In this paper, a tomato maturity dataset is constructed, which contains 1812 images and 4000 targets. The results showed that the YOLOv5-AT2 model proposed in this paper mAP was 88.06%. The detection average precision (AP) for green tomatoes and red tomatoes was 86.12% and 90%, respectively. The detection accuracy of green tomato and red tomato were 81.25% and 92.77%. This method can achieve fast detection of tomato maturity in complex environments and provides a new idea for tomato detection.