基于深度卷积神经网络的柑橘目标识别方法

Detection Method of Citrus Based on Deep Convolution Neural Network

针对户外自然环境,基于深度卷积神经网络设计了对光照变化、亮度不匀、前背景相似、果实及枝叶相互遮挡、阴影覆盖等自然环境下典型干扰因素具有良好鲁棒性的柑橘视觉识别模型。模型包括可稳定提取自然环境下柑橘目标视觉特征的深层卷积网络结构、可提取高层语义特征来获取柑橘特征图的深层池化结构和基于非极大值抑制方法的柑橘目标位置预测结构,并基于迁移学习完成了柑橘目标识别模型训练。本文运用多重分割的方法提高了柑橘目标识别模型的多尺度图像检测能力和实时性,利用包含多种干扰因素的自然环境下柑橘目标数据集测试,结果表明,柑橘识别模型对自然采摘环境下常见干扰因素及其叠加具有良好的鲁棒性和实时性,识别平均准确率均值为86. 6%,平均损失为7. 7,平均单帧图像检测时间为80 ms。

Citrus detection and location is the foundation of citrus automated picking systems, in light of the outdoor natural picking environment, a citrus visual feature recognition model was designed based on deep convolution neural network with good robustness for typical interfering factors, such as illumination change, uneven brightness, similar foreground and background, mutual occlusion of fruit, branches and leaves, shadow coverage and so on. The model included a deep convolutional network structure which can steadily extract the visual features of citrus under natural environment, a deep pool structure which can extract high-level semantic features to get citrus feature map, a citrus location prediction model based on non-maximum suppression method. Moreover, the proposed model was trained by transfer learning method. Each raw image was segmented into several sub-images before citrus detection to enhance the ability of multi-scale object detection, and reduce the computing time of citrus detection. A testing dataset, which contained representative interference factors of natural environment, was used to test the citrus detection model, and the proposed detection model had good robustness and real-time performance. The average detection accuracy and the average loss value of the model was 86.6% and 7.7, respectively, meanwhile, the average computing time for detecting citrus from single image was 80 ms. The citrus detecting model constructed by deep convolution neural network was suitable for the citrus harvesting in the natural environment.