The Research Of Farmland-weed Identification Algorithm And Its Implementation On The DSP

The environment adaptability of the weeds in the farmland is stronger than the crops', so more moisture, nutrients and growing space are depredated by the weeds. As a result, the normal growth of crops is severely disrupted and the yield is reduced. The quality of the fruit from the crops is declined. According to results reported by the Food and Agriculture Organization of the United Nations (FAO), there are about 8000 kinds of weeds in the farmland all over the world and about 250 kinds of them disturb the growth of the crops. Every year, more the 95 billion dollars are lost as a result of the harm caused by the weeds. Therefore, how to effectively remove the weeds has become a concern of many experts and scholars and much research has be conducted.Nowadays, there are many ways to eliminate the weeds. Among of them, the technique that sprays the herbicide accurately is widely used with the advantages of low cost, low pollution and easy operation. The key technology of precise spraying is to estimate the coverage and the location of the weeds with machine vision. With the development of embedded technology, the performance of the microprocessors has been greatly improved. Some high-performance DSP (Digital Signal Processor) can work with the core-clock frequency up to 1000MHz or even higher. The unique core mechanism can achieve parallel executions of multiple instructions, so its processing capabilities are greatly enhanced. Based on the background, this paper do some research on the identification algorithms for the weeds in the farmland of drill and implant them on the DSP in order to achieve precision-herbicide application.In this paper, the characters of color and location of the plant in the farmland are adopted to achieve the research of weed-identification technique.In order to extract the plant from the image, the color image is transformed into the gray one to enhance the plant and weak the background. In the color space of RGB, five combinations of the R, G and B color components are researched. They are 2G-RB, (G-R)/(G+R), 3G-2.4R-B, improved 2G-R-B and improved 3G-2.4R-B. The enhanced effects of them are compared. The research proves that the method of improved 3G-2.4R-B works with the highest rate of correct segmentation and the most stable accuracy in the five of them. However, it is the most time-consuming one. The combination method of 2G-RB works with the middle performance about the correct segmentation rate and the stability in the five methods and it consumes the least time.According to the characteristic of the histogram for the gray image, a threshold is used to the change the gray image into the binary one and segment the plant from the background. The test result of 35 images shows that the thresholds determined by the iteration method and the maximum variance between the classes are almost the same and there is nearly no difference in the effect of the segmentation. But the time efficiency is different. The iterative method can save about 84.34% time that consumed by the maximum variance one. With a reasonable initial value, the times of the iteration can be reduced and more time will be saved.By taking the advantage of the location character of the drill, the crop can be removed from the combination image of crops and weeds. The drills distribute in line and the line space is fixed. In the binary image, there are more non-zero pixels of the crop lines than the non-crop lines. So, by summing the non-zero pixels in the crop-line direction, the center of the crop lines can be determined according to the peak values. In order to filter the crop, five methods are studied. They are the fixed-width algorithm, the projection algorithm, the window algorithm, the labeling algorithm and the improved labeling algorithm. All of them are tested and analyzed. The result gives out that the improved labeling algorithm is able to removed the crop effectively and at the mean time retain more information between the crop rows. So, more information about the weed coverage can be provided. In order to make the crop-filtering algorithm more adaptable, a method that combines the projection algorithm and the improved labeling algorithm is presented in this paper. The two algorithms are selected automatically depending on the aggression degree of the weeds.By counting the weed pixels remained in the image, the statistic information of weed coverage can be obtained. It will be used to guide the herbicide-spraying device. And the locations of the weeds can be gained by calculating the centroid of the weed pixels.In order to facilitate the algorithm development on the DSP, a platform with the combination of MATLAB and CCS (Code Composer Studio) is built. With the help of the interface functions provided by CCSLink, MATLAB can control and monitor the running state of the DSP. And the data transmission is implemented in real time through the RTDX (Real Time and Data Exchange) technology. The weed-identification algorithms are achieved on the DSP by merging the C language, the inline functions and the assembly language. The optimization function provide by the compiler in CCS is used. Considering the features of the instructions and the software pipelining mechanism in the C64x+ core, some techniques, such as loop unrolling, data packing and so on, are utilized to optimize the codes. After the optimization of the code implementations, such as the histogram algorithm, the graying algorithm, the median-filtering algorithm, the labeling algorithm and so on, the performance of the codes are greatly improved. The efficiency of the codes can be increased by 73.69%. The testing data shows that the weed-identification algorithm implemented on the DSP can complete the identification of a image with 320*240 pixels between 14.935731ms and 18.865319ms. With the performance, at least 50 images can be identified in one second, which meets the real-time application.