基于高光谱图像与光谱特征融合技术的鸡蛋新鲜度无损判别模型的建立

Establishment of Non-Destructive Discriminant Model for Eggs Freshness Grade Based on Fusion Technology of Image and Spectral Features of Hyperspectral

摘要:

鸡蛋新鲜度等级评价是鸡蛋品质检测过程中的一项重要技术指标。选取了不同储藏环境的鸡蛋样本并采集其高光谱图像信息与光谱信息,提取图像特征和光谱特征；采用并行式融合方法进行图谱特征融合,基于连续投影法-灰度共生矩阵方法进行特征提取；建立支持向量机鸡蛋新鲜度判别模型。采用粒子群算法优化模型,训练集准确率达到85%,预测集准确率达到76.67%。为了解决单模型可能出现的偶然性误判问题,采用递进式特征融合方法,引入多模型共识策略和深度残差网络ResNet 50分析方法。建立基于连续投影法-方向梯度直方图特征提取方法的多模型共识策略,该模型的训练集准确率提升至89%,预测集准确率提升至88%；同时,建立基于连续投影法-方向梯度直方图特征提取方法的深度残差网络ResNet 50模型,模型的训练集准确率提升至89%,预测集的准确率提升至86.67%。图谱特征融合建模分析表明,并行式融合方法和递进式融合方法对鸡蛋新鲜度等级判别都有一定的可识别性,且递进式融合算法的多模型共识策略判别效果更佳。

Abstract:

Egg freshness grade evaluation is an important technical indicator in process of egg quality inspection. Egg samples from different storage environment were prepared, the hyperspectral image information and spectral information of eggs were collected, and the image features and spectral features were extracted. The image features and spectral features were integrated with parallel integration method, and the features were extracted based on successive projections algorithm and gray-level co-occurrence matrix method. The support vector machine egg freshness discriminant model was built. The model was optimized by particle swarm optimization algorithm, the accuracy rate of training set reached up to 85%, and the accuracy rate of prediction set reached up to 76.67%. In order to solve the occasional misjudgment of single model, the progressive features integration method was used, and the multi-model consensus strategy and deep residual network ResNet 50 analysis method were introduced. The multi-model consensus strategy based on successive projections algorithm-histogram of oriented gradients features extraction method was built, the accuracy rate of training set of the model increased to 89%, and the accuracy rate of prediction set increased to 88%. Meanwhile, the deep residual network ResNet 50 model based on successive projections algorithm-histogram of oriented gradients features extraction method was built, the accuracy rate of training set of the model increased to 89%, and the accuracy rate of prediction set increased to 86.67%. The image features and spectral features integration modelling analysis indicated that both parallel integration method and progressive integration method had a certain identifiability for egg freshness grade discrimination, and the multi-model consensus strategy of progressive integration method showed better discrimination effect.