Research on tobacco growth evaluation algorithm based on improved DeepLabv3+

The growth and development of tobacco plants are crucial for determining both the quality and yield of tobacco. Canopy coverage area serves as a significant indicator of tobacco plant growth, effectively reflecting the growth status of tobacco plants. Traditional methods for assessing plant growth often suffered from low efficiency, strong subjectivity, and cumbersome operations. This research employed deep learning models for precise semantic segmentation of tobacco plants and proposed an improved DeepLabv3+ method for canopy semantic segmentation to evaluate plant growth. The method utilized MobileNetV2 as the backbone network for feature extraction, aiming to streamline model parameters and reduce computational burden. Furthermore, ODConv was introduced to replace atrous convolution, mitigating the loss of critical feature information caused by occlusion between tobacco plants and leaves. Finally, a triplet attention mechanism was employed to enable the model to more accurately distinguish between tobacco plants and background, enhancing the recognition of key features and detail preservation. The results demonstrated that this method effectively differentiated tobacco plants from the background, achieving an mIoU of 85.38%, mPA of 91.65%, and Accuracy of 92.28%. Compared to the original DeepLabv3+ model, the mIoU, mPA, and Accuracy were improved by 2.17%, 1.45%, and 1.22%, respectively. This approach provided reliable technical support for subsequent evaluation of tobacco plant growth.