提出了一种新的直接从图像中检测特定目标的算法。算法分为3个步骤。首先用分水岭变换对图像作“过分割”，得到标注的积水盆地和堤坝。实际目标可能是其中一个或多个盆地的组合。然后对盆地进行初步筛选，删除不可能是目标的盆地，并用区域邻接图（RAG）表示剩下的盆地。最后针对问题特点定义能量函数，提出一种启发式遗传算法，用来在RAG中检测使得能量函数极小化的子图，子图对应的区域就是目标。实验结果表明了该算法的有效性。

In this paper, a new method for detecting specific objects in the image is presented. It consists of three steps. Firstly, a watershed transformation (WT) is used for over segmenting the image in different small basins and dams. The desired object should be one of these basins or the combination of some ones. Secondly, the basins which are most impossible to be the objects are deleted, and then the neighboring relationships between the remainders are analyzed to construct a region adjacency graph (RAG). Finally, an energy function is constructed and a heuristic genetic algorithm (GA) is proposed to extract the optimized sub graphs from the RAG. The regions corresponding to the sub graphs are the desired objects. Experiment results demonstrate the feasibility of this approach.