毛桃从果树上成熟到最后到达消费者手中, 中间需要经过采摘、 包装、 运输等一系列过程, 在每一个过程中都有可能产生碰伤果。 因此查看哪一个过程产生的碰伤果最多, 从而对加工过程进行针对性地改进就显得尤为重要。 纵观国内外高光谱技术在检测水果碰伤方面的应用, 绝大多数都是忽略图像特征而只使用了光谱特征, 基于图像特征结合光谱特征建模的少之又少。 其次在水果碰伤时间定性判别方面, 多以天数为间隔, 时间间隔较大意味着水果碰伤时间越久, 其变化越明显, 检测准确率也就越高, 目前尚缺乏有效方法对于碰伤时间较短的水果进行碰伤时间分类。 以90个模拟表面碰伤的毛桃为实验样本, 分别采集毛桃碰伤12, 24, 36和48 h后的高光谱图像。 毛桃样品的光谱特征提取是采用感兴趣区域的100个像素点的平均光谱以防止单个像素点的光谱信息与整体光谱信息差距较大; 通过主成分分析(PCA)对毛桃图像进行降维后选取最能体现毛桃碰伤的PC1图像, 在 PC1图像的权重系数曲线中波峰波谷处挑选出4个特征波长点(512, 571, 693和853 nm)作为特征图像, 特征图像灰度化操作后计算得到平均灰度值作为毛桃碰伤图像特征。 最后基于最小二乘支持向量机(LS-SVM)算法分别建立毛桃碰伤时间的光谱特征模型、 图像特征模型以及图像特征结合光谱特征模型共三种判别模型, 并且根据其分类准确率来判断模型的性能。 结果表明: 三种毛桃碰伤模型的分类准确率都随碰伤时间的增加而增加; 基于径向基核函数(RBF_kernel)建立的图像特征结合光谱特征的模型预测效果最好, 对碰伤12, 24, 36和48 h的毛桃样品识别正确率分别为83.33%, 96.67%, 100%和100%, 这可能是由于具有非线性特点的径向基核函数所建立的模型更加适合用于毛桃碰伤时间的分类。 图像特征结合光谱特征的模型能够较好地实现对水果碰伤时间的估计, 可为水果外部品质分选提供一定的参考和依据, 并对水果销售和深加工企业具有一定的借鉴意义。

From the ripening of the fruit tree to reaching the consumers, the peaches need to go through a series of processes such as picking, packaging, and transportation. In each process, bruised fruit may occur. Therefore, it is particularly important to check which process produces the most bruises and to improve the processing process in a targeted manner. Throughout the application of hyperspectral technology in detecting fruit bumps at home and abroad, most of them ignore image features and only use spectral features. Modeling based on image features combined with spectral features is rare. Secondly, the interval is usually the number of days in terms of the qualitative judgment of fruit bump time. The larger time interval, the longer fruit bump time, and the more obvious change, the higher detection accuracy. There is no effective method of classifying the bump time for the fruits which were bruised in a very short time. In this paper, 90 simulated surface bruises were taken as experimental samples, and hyperspectral images of the bruises 12, 24, 36 and 48 h were collected respectively. The spectral feature extraction of the peach sample uses the average spectrum of 100 pixels in the region of interest to prevent the spectral information of a single-pixel from being significantly different from the overall spectral information; The PC1 image that can best reflect the bruise of the peach is selected after dimensionality reduction by principal component analysis (PCA). In the weight coefficient curve of the PC1 image, 4 characteristic wavelength points (512, 571, 693, 853 nm) at the peak and valley points are selected as the characteristic image. The average gray value which calculates as the characteristic image after graying is used as the feature of the bruised peach image. Finally, based on the least squares support vector machine (LS-SVM) algorithm, three discriminant models, namely the spectral feature model, image feature model and image feature combined with the spectral feature model of the peach bruise time were established, and the performance of models was judged according to their classification accuracy. The research results show that the classification accuracy of the three peach bruise models increases with the increase of bruise time; the model based on the radial basis kernel function (RBF_kernel) combined with the spectral features has the best predictive effect, and it has the best prediction effect on bruises. The recognition accuracy rates of the peach samples at 12, 24, 36 and 48 h were 83.33%, 96.67%, 100% and 100%, respectively. This may be due to the model established by the radial basis kernel function with nonlinear characteristics is more suitable for peach Classification of bump time. The model combining image features with spectral features can better estimate the fruit bump time, and it can provide a certain reference and basis for fruit external quality sorting, which has certain reference significance for fruit sales and deep processing enterprises.