基于改进型SegNet的苹果采摘点分割算法研究

Apple picking point segmentation based on improved SegNet

针对直接使用SegNet模型处理苹果图像会出现采摘点分割不清晰和分割错误的问题,提出了一种改进的SegNet语义分割模型进行苹果采摘点分割,使其更适应于复杂的自然环境,为采摘机器人提供帮助。在SegNet模型中引入DenseNet的思想,直接连接来自不同网络层的特征图,实现图像特征的多次重用,以提高模型的分割精度。为了验证改进算法的有效性,选取3种不同品种的苹果建立图像数据集,并在PyTorch深度学习框架上进行训练。利用通用的评价指标,将SegNet模型改进前后的测试结果进行对比。试验结果表明,改进的SegNet模型的最佳精确率、召回率、特异性和Dice系数分别为83.10%、84.82%、98.56%和83.95%。相比原模型,改进的SegNet模型识别成功率提高了2.19%,在运算时间几乎不变的情况下,能够更好地实现自然环境下采摘点的分割,为其他种类水果的采摘点分割识别算法提供了研究基础。

Aiming at the problems of unclear segmentation and wrong segmentation of picking points when using SegNet model to process apple image directly, an improved SegNet semantic segmentation model was proposed to segment apple picking points, which made it more suitable for complex natural environment and provided help for picking robot. The idea of DenseNet is introduced into SegNet model, which directly connects feature graphs from different network layers to achieve multiple reuse of image features, so as to improve the segmentation accuracy of the model. To verify the effectiveness of the improved algorithm, three apples of different varieties were selected to build image data sets and trained on PyTorch deep learning framework. The test results of SegNet model before and after improvement are compared with the general evaluation indexes. The experimental results show that Precision, Recall, Specificity and Dice coefficients of the improved SegNet model are 83.10%, 84.82%, 98.56% and 83.95%, respectively. Compared with the original model, the improved SegNet model improved the recognition success rate by 2.19%. Under the condition of almost constant computing time, the SegNet model could better realize the segmentation of picking points in the natural environment, providing a research basis for the segmentation and recognition algorithm of picking points of other kinds of fruits.