基于视觉注意模型和最大熵分割算法, 提出了一种自适应显著目标分割方法来分离目标和复杂背景, 以便快速准确地从场景图像中检测出显著目标。首先, 通过颜色、强度、方向和局部能量4个特征通道获取图像的显著图; 通过引入局部能量通道来更好地描述了显著目标的轮廓。然后, 根据显著图中像素灰度的强弱构建不同的目标检测蒙板, 将每个蒙板作用于原图像作为预分割的结果, 再计算每个预分割图像的熵。最后, 利用最大熵准则估计图像目标熵,根据预分割图像的熵和目标熵判断选取最优显著目标分割图像。实验结果表明: 本文算法检测的显著目标更为完整, 分割性能F-measure达到0.56, 查全率和查准率分别为0.69和0.41, 相对于传统方法更为有效准确, 实现了在复杂背景下对显著目标的有效准确检测。

On the basis of a visual attention model and a maximum entropy segmentation method, an adaptive segmentation method was proposed to segment the object from a complex background in the scene image and to detect a salient object effectively and accurately. First, the feature of original image was extracted via four channels on color, intensity, orientation and local energy. The profile of object feature was described more accurately by combining the channel of local energy with a simple biologically-inspired model. Then, object detection masks were constructed to remove background gradually according to the gray intensity of the pixels in the saliency map. By taking blend masks with the original image as a pre-segmentation result, the entropy of pre-segmentation images was computed. Finally, the entropy of salient object was estimated via maximization information entropy principle and the optimized image extraction for the salient object was obtained by estimating the relationship of entropy between salient object and masks in the saliency map. Experimental results indicate that the salient object detected by proposed method is more integrity, the F-measure of segmentation performance is 0.56, and the precision ratio and the recall ratio of detection are 0.69 and 0.41, respectively. The proposed method is more reasonable and effective than the traditional method, and it can satisfy the requirements of detecting the salient objects from complex backgrounds.