Automatic recognition of corn straw coverage based on fast Fourier transform and SVM

The crop residues auto-detection is an important technology for intelligent agriculture. However, the farmland condition is complicated that it’s difficult to detect the crop residues and calculate its coverage. The field is always unevenly distributed with clays and gullies, besides that, the crop residues have various shapes and sizes. Those variables affect the detection accuracy. The research shows that FFT (fast Fourier transformation) can distinguish both the high frequency and low frequency signals in the image. The high frequency signals characterize the outline of the image and the low frequency signals characterize the content of the image. The crop residues belong to high frequency information exactly. Meanwhile, SVM (support vector machine) has a high distinguishing ability for target recognition in complex backgrounds. It maps linearly indivisible data to high dimensional space through different kernel functions and converts it into linearly separable data. It establishes a maximum interval hyperplane in high dimensional space and two mutually parallel hyperplanes on both sides of the data hyperplane. On the basis above, this paper proposed a novel fusion algorithm for crop residues auto-detection with Fast Fourier Transformation (FFT) and Support Vector Machine (SVM). We also designed and realized an intelligent monitoring equipment for crop coverage auto-detection which consists camera, embedded board, location module, communication module and other sensors through 4G cloud network. In order to detect residues, we extracted the high frequency information (residues region) by designing high pass filter after FFT with different kernels. We tested three different filter kernels including square, circle and diamond filter with three shapes. The results of the experiments indicated that the square filter with size 11×11 was the best choice considering speed and accuracy. Then, an adaptive threshold segmentation method was put forward to process the normalization image we obtained. The threshold was set by calculate the energy score we defined based on the idea of integral. The experiments showed that 0.8 was the best threshold choice for our algorithm. It can preserve the most crop region and remove the un-crop region as much as possible. We used the segmentation result as the detection mask for the next step. On the other hand, we trained a SVM classifier using our database with different input sizes and kernels and selected the parameter group with best performance as final choice. The classifier was utilized to re-classify the adaptive threshold segmentation result. Finally, we found that the traditional image erosion and dilation algorithm treated all pixels without considering its credibility. So it was prone to misprocessing (including etching away areas that should be preserved, or expanding the noise area that should be removed). To avoid this misprocessing, we raised a multi-scale filter to erase noises and fill tiny holes which was similar to image pyramid. There were two filters with size 3×3 and 7×7. The bigger one was to filter the image noise with tinny size and the smaller one was to expand the residues region with holes. We benchmarked the algorithm on our database and the results indicated that our method was state of the art. The error was about 4.55% and it just used 0.05 s.