Study On Weed Identification In Cabbage Field

Weeding is an important part of agricultural production,nowadays,the main methods of which are cultivating weeding and chemical pesticides and weeding.The former is inefficient and has a large workload while the latter is highly harmful.With the development of mechanical and artificial intelligence,weeding robots have gradually entered people's horizon.However,the existing weeding robot's visual recognition systems are generally bulky,and the overall price is relatively expensive.Based on the shortcomings of traditional weeding robots,this paper proposes to use smart phones as the visual recognition system in this experiment.The system has the functions of image acquisition and processing,and has the characteristics of small size,low power consumption and no occupation of the robot body space.Currently,the price of the medium-performance mobile phone is not too high,which reduces the overall cost of the weeding robot so that a simple APP with image processing function was developed in this experimental.In the traditional image recognition process,the acquired image is generally preprocessed in one step or several steps and then segmented.The skeleton data of the segmented target object is extracted and the object category is finally judged according to the mode identification method.In this paper,RGB color images are converted to HSV and L*a*b* color spaces,and gray images are acquired using single channel separation.The a* channel image was selected by contrast analysis for Gaussian filter denoising.Furthermore,the fixed threshold segmentation algorithm and the OTSU threshold segmentation algorithm are used to segment the grayscale image.Comparing the segmentation results,the OTSU threshold segmentation algorithm is better in segmentation under natural illumination and has better adaptability.Finally,the area is selected as the identification feature based on the characteristics of cabbage and weeds in the same production period,and 41% recognition accuracy rate in 100 verification images was abtained.In the field of instance segmentation of deep learning,Mask R-CNN algorithm achieves good segmentation effect in urban environment so that the algorithm is used to identify weed seedlings and cabbage seedlings under natural light in this paper.Three common weed seedlings and cabbage seedlings were selected as the training set training network,and the pass rate of the test set was 67%.Under the condition of qualified identification,the algorithm shows a good generalization ability taht the objects in the image can be accurately identified under different illumination and background,which solves the problem that the traditional image recognition algorithm is difficult to segment in the natural environment and the complicated classifier setup work is avoided at the same time.