自然环境下重叠果实图像识别算法与试验

针对非结构化自然环境中光照变化和对象重叠特征等外界因素给图像处理带来的难题,提出了一种自然环境下重叠果实的图像识别与边界分割的组合优化算法。该组合优化算法首先对原始图像进行噪声滤波处理,然后利用Sobel算子以及改进算子的最大类方差法(OTSU)来辨识重叠果实目标;接着采用K-means算法对重叠目标的像素进行聚类得到单个目标位置,再结合边缘检测结果的连通域分析及区域生长获得单个目标边界的大致区域;最后利用基于限制区域的分水岭算法,得到目标的精确边界。为了验证所提算法的有效性和适应性,进行了试验研究。试验结果表明:所提出的组合优化算法不仅可以在自然环境下从重叠物体图像背景中识别出重叠目标,而且还可以从重叠目标中分割出单个目标的精确边界。

Image Recognition Algorithm and Experiment of Overlapped Fruits in Natural Environment

A combined algorithm for image recognition and boundary segmentation of overlapped objectives under natural and unstructured environments was proposed. The algorithm dealt with the challenging problem of image processing applied to the agriculture with complicated external factors, such as illumination changes in unstructured natural environment, objective feature overlapping. Firstly, the image noise on the original image received from the camera was filtered out using bilateral algorithm. Secondly, overlapped objectives in the filtered image were recognized by OTSU algorithm based on the improved operator. Then the single object position was obtained by using K-means clustering algorithm on the pixels of the overlapped objectives. Afterwards using Sobel operator or Canny operator, the approximate area of the single object was recognized by connected domain analysis on the edge detection results and the domain growth. Finally, after internal and external reception basins received from the area of the object, the position of the single object boundary was confirmed, the precise contour of the single object was obtained by using watershed algorithm on the restricted area which was the area between internal and external reception basins. In order to verify the effectiveness and applicability of the proposed algorithm, several experiments were carried out, and only two experiments were shown due to the limited space. The first experiment chose a relatively simple image with overlapped tomatoes under the simple image composition, the second experiment chose a complicated image to further verify the adaptability of the algorithm. The experimental results showed that the proposed algorithm can recognize the overlapped objectives under natural environments and it can also segment the single object from the overlapped objectives.