复杂背景下的红外图像往往由于噪声较多、背景区域重叠、目标与背景对比度较差等因素, 在对目标区域分割时会造成过分割或欠分割。针对此现象, 提出了一种将全卷积神经网络和动态自适应区域生长法相结合的红外分割算法。首先利用全卷积神经网络对目标区域在像素级别进行特征提取, 通过神经网络强大的自学习能力获得目标区域的粗分割结果; 然后根据粗分割结果, 对其取外接最小面积矩形框, 并根据矩形框位置在原始图像上确定目标区域, 并以此矩形区域进行动态自适应区域生长, 形成第二次分割结果。最后融合全卷积网络(FCN)的粗分割结果和区域生长分割结果, 实现目标区域的最终分割和提取。仿真实验表明, 该方法能有效利用FCN对红外图像复杂背景的消除能力, 而区域生长法对分割细节的敏感也同时弥补了FCN分割精度的不足, 取得了较好的分割效果。

Infrared images in complex backgrounds tend to be over-segmented or under-segmented when segmenting the target region due to factors such as high noise, overlapping background regions, and poor target and background contrast. Aiming at this phenomenon, it proposes an infrared segmentation algorithm combining the full convolutional neural network and the dynamic adaptive region growing method. Firstly, the full-convolution neural network is used to extract the feature of the target region at the pixel level, and the coarse segmentation result of the target region is obtained by the powerful self-learning ability of the neural network. Then, according to the result of the coarse segmentation, the minimum area rectangular frame is taken outside, and according to the position of the rectangular frame, the target area is determined on the original image, and the dynamic adaptive area growth is performed by using the rectangular area to form a second segmentation result. Finally, the rough segmentation results and regional growth results of the FCN are combined to achieve the final segmentation and extraction of the target region. The simulation experiments show that the method can effectively utilize the FCN to eliminate the complex background of infrared images, and the sensitivity of the region growing method to the segmentation details also compensates for the lack of FCN segmentation accuracy and achieves a good segmentation effect.