基于模糊集理论的苹果表面阴影去除方法

为了提高阴影影响下的苹果目标提取精度,该文提出了一种基于模糊集理论的苹果表面阴影去除方法。该方法将含阴影图像作为一个模糊矩阵,利用所设计的隶属函数进行图像去模糊化处理,达到图像增强的目的,进而削弱苹果表面阴影对目标分割的影响。为了验证算法的有效性,采用基于灰度阈值和基于颜色聚类2种算法对去除阴影前后的目标图像进行分割,并选用分割误差、假阳性率、假阴性率和重叠系数4项指标进行了分析比较,试验结果表明,去除阴影之后,2种分割算法所提取的苹果目标区域较去除阴影之前有了较大的提高,2种分割算法的平均分割误差分别为3.08%和3.46%,比去除阴影之前降低了20.53%和25.92%,假阳性率、假阴性率分别降低了29.79%、29.98%和21.25%、29.83%,重叠系数分别提高30.96%和24.55%。与灰度变换法去除阴影后分割的效果比较表明,该方法的平均分割误差降低了29.23%,假阳性率、假阴性率分别降低了30.97%和20.40%,重叠系数提高了26.60%;与直方图均衡化法的比较表明,分割误差降低了25.59%,假阳性率、假阴性率分别降低了22.74%和27.56%,而重叠系数提高了27.43%。这一系列数据表明,基于模糊集理论的阴影去除方法具有较好的阴影去除效果。经过去除阴影后,可以获得更高的目标分割性能,目标提取精度显著提高,表明将模糊集方法应用于苹果目标的阴影去除可以有效地提高苹果目标区域的提取精度。

Shadow removal method of apples based on fuzzy set theory

Abstract: Illumination changes and shadows are the problems that must be considered in the recognition of fruits in nature scenes. In order to improve the accuracy of apple extraction under the influence of shadows, an apple surface shadow removal algorithm based on a fuzzy set was presented. In this algorithm, the image including shadows could be seen as a fuzzy matrix. The membership function was used for image de-blurring processing so as to enhance the image and then weaken the shadow's influence on apple segmentation. After the usage of a fuzzy set theory, the saturation of each pixel should be enhanced so as to reduce the difference of the adjacent pixel points. In order to verify the validity of the algorithm, a gray threshold algorithm and k-means color clustering algorithm were adapted to segment targets before and after shadow removal. In addition, the gray threshold method used in this paper was an Otsu adaptive threshold which could be objective for the judgment of segmentation results. For the k-means method used in this paper, the parameter k was selected as k=3, which means that all the images were clustered into leaves, branches, and apples. In this paper, four criteria such as Af (Segmentation error), FPR (False Positive Rate), FNR (False Negative Rate), and OI (Overlap Index) were used to evaluate the segmentations results. The results showed that after shadow removal, the target extraction area using the two segmentation algorithms were larger than that before shadow removal. The average segmentation error was 3.08% and 3.46% of the two segmentation algorithms, and it decreased 20.53% and 25.92% respectively when compared with the result before shadow removal. The FPR and FNR decreased 29.79%, 28.98%, 21.25%, and 29.83%, and OI increased 30.96% and 24.55%. In order to further verify the validity of this algorithm, the proposed algorithm was compared with a gray scale transform method and histogram equalization method. The experimental results showed that under the method proposed in this paper, the average segmentation error was reduced by 29.23% compared with the result obtained by the gray scale transform method, FPR and FNR was reduced by 30.97% and 20.40%, while OI increased by 26.60%. Then the method proposed by this paper was compared with the histogram equalization method, resulting in the segmentation errors being reduced by 25.59%, and FPR and FNR was reduced by 22.74% and 27.56%, while OI increased by 27.43%. This series of data showed that the presented algorithm could get better shadow removal effects. All these results showed that the presented shadow removal algorithm proposed by this paper could improve the target segmentation performance and are feasible and effective to remove the shadows of apples. This method has very broad further significance.