为了更好地实现不同光照条件下的农作物图像融合，在Contourlet变换（contourlet transform, CT）的基础上采用了适合农作物图像的融合规则进行了融合处理。首先，采用 Contourlet 变换对源图像进行多尺度、多方向分解，得到低频子带系数和带通方向子带系数。然后，针对低频子带系数的选择，采用了一种改进的线性加权融合方法，以期减小噪声对融合结果的影响；针对带通方向子带系数的选择，结合人眼视觉特性，采用了一种基于梯度最大化规则的系数选择方案，得到待融合图像的系数。最后，经过 Contourlet 逆变换得到融合图像。与小波变换方法（wavelet transform, WT）进行了融合结果的比较，结果表明，与WT方法相比，该文方法在互信息量（mutual information, MI）、空间频率（spatial frequency, SF）、均方差（mean square error, MSE）、信息熵（entropy， Ent）、相关系数（correlation coefficient, CC）、平均梯度（average gradient, G’）和峰值信噪比（peak signal to noise ratio, PSNR）指标上均有了较大提升，表明利用该方法可以取得优于WT的融合效果；在此基础上，利用CT常见融合规则与文中融合规则进行了比较，同样表明 CT 方法可以有效提高图像融合的效果。研究表明，将文中所采用的融合规则应用于不同光照条件下的农作物图像融合是有效的、可行的。该研究可为变光照条件下的作物图像融合技术提供参考。%In the research field of agricultural crop growth state monitoring, it is difficult to capture an image that can meet the needed usage for a whole description of a crop’s growth information. Image fusion could combine two or more source images into a single composite image with extended information, and it is very useful for a crop’s monitoring system. Aiming at solving the fusion problem of crop images in different light conditions, a proper image fusion algorithm based on contourlet transformation (CT) theory was carried out. First, by using CT, all the source images were decomposed into multi-scale and multi-direction sub-bands; then, for the low frequency coefficients, linear weighted fusion rules were adopted to reduce the influence of noise. For the band-pass directional sub-band coefficients, a maximum gradient rule was used to meet the human visual characteristics. Finally, the inverse transformation of contourlet was used to get the fused image. To testify as to the performance of the algorithm, cucumbers, cherry tomatoes, eggplant, and peppers captured in different light conditions were used as experimental crops. The experiment was divided into two parts: (1) a performance comparison between CT and WT, (2) a comparison of different fusion results of CT. A comparative test by using WT method showed that the proposed method could get much better performance than wavelet based fusion methods and commonly used fusion rules. In particular, the MI of fused image was 27.04% higher than the wavelet based method, SF value increased by 37.73%, MSE parameter was 46.97% higher, Ent was also improved by 19.69%, CC had a little enhancement by 2.76%, and G’ was 11.21% higher than the WT based image fusion algorithm. Also, PSNR value was boosted by an average 8.06%. Another comparative experiment with commonly used fusion rules showed that edge strong information was 0.30% higher, structure similarity was promoted by 0.50%, and the average gradient was boosted by 2.63% more than under “linear+max” rules, and entropy value was 5.07% higher than under “min+mean” rules. These experiments all showed that the proposed CT based image fusion algorithm was practical and valid for agricultural product image fusion in different light conditions. This research provides a useful reference for the fusion of crops in different conditions.
展开▼
