基于高光谱技术的菌落图像分割与计数

在平板菌落计数过程中，菌落与背景区域类似的颜色会干扰菌落的准确计数。为了准确测定细菌数，该研究利用高光谱图像技术捕捉成分差异引起的菌落与背景区域光谱特征，并结合化学计量学方法对平板的菌落进行分割并实现计数。采集枯草芽孢杆菌菌落平板的高光谱图像，提取菌落、背景区域的高光谱信息；利用遗传算法结合最小二乘支持向量机建立菌落区域/背景区域判别模型；随后，将菌落平板高光谱图像中每一个像素点对应的光谱信息代入判别模型以判断属于菌落的区域，模型的识别率为97.22%；最后，利用特征波段下的高光谱图像实现菌落的分割及计数，计数平均相对误差值为4.2 %，用时约为10 min。相比较于计算机视觉计数法，菌落计数法的平均相对误差降低了49.4%，结果表明建立的方法有望成为一类新的准确平板菌落计数方法。

Colony image segmentation and counting based on hyperspectral technology

Abstract: Colony plate counting methods as the national standard method are common and traditional for quantity inspection of living bacteria in food and agricultural products. The colonies are counted by manual counting and computer vision counting methods because of the color difference between colonies and background (such as medium and the edge of petri dish). However, colonies and background with similar color will interfere with the colony segmentation and lead to the deviation in counting results. Considering colonies and background have different spectral information resulting from different chemical compositions, hyperspectral imaging technology can be applied to colony segmentation and adherent colonies separation combined with chemostics. The segmentation method of colony and background, separated method of adherent colonies and calculation method of colony number were developed for colony counting. The hyperspectral images of Bacillus subtilis (B. subtilis) colony plate were acquired in the wavelengths from 431 nm to 963 nm. Spectral information of colonies and background (medium and the edge of petri dish) was extracted after preprocessing hyperspectral images. Genetic algorithms (GA) was used for processing spectral data and eleven characteristic wavelengths were selected (604, 636, 790, 799, 748, 683, 492, 437, 558, 470 and 928 nm). GA-least square support vector machine (GA-LS-SVM) model was established for distinguishing colonies and background by using the spectral information at the eleven characteristic wavelengths. The identification model with identification rate of 97.22% indicated that the colony and background could be successfully distinguished. The segmentation method of colony and background was developed. The spectral information of every pixel was extracted to identify whether it is colony or background by using the identification model. The binary image of colony segmentation was obtained through the spatial information in hyperspectral images. The location of colony was assigned as 1 and the location of background was assigned as 0, resulting in colony segmentation. The hyperspectral image at 604 nm was used for segmentation of adherent colonies to obtain binary image of adherent colonies segmentation. The segmentation threshold between background and colony was set as 0.5. The results demonstrated that the colonies were successfully segmented from background and adherent colonies could be accurately separated. Finally, the isolated colonies were counted by contour tracking algorithm after colony segmentation from background and adherent colonies separation. For application, acquisition of hyperspectral image, colony segmentation, adhesion colonies separation and colony number calculation were used for B. subtilis colony counting of real samples. The time required by the developed method was about 10 mins. Its average relative error of colony count was 4.2% with the manual counting method as the standard method. In addition, the colony plate counting was performed by using computer vision counting method. The average relative error of colony counting was 8.3% which was higher than the developed method. These results indicated that this method performed better than computer vision counting method though the consuming time was longer than that spent by the automatic colony counter. This method with high accuracy can become a novel plate colony counting method and provide technical support for the detection of microbial quantity in food and agricultural products.