关键词: 作物与杂草识别, 深度学习, 卷积神经网络, 逐像素分类, 语义分割

Abstract: 【Objective】In order to improve the accuracy and real time recognition of crops and weeds, the field color image of seedling beet was taken as the research object, and a pixel-wise classification method based on deep separable convolution was proposed.【Method】In this paper, the field color image of the seedling beet was used, the pixels in the color image were tagged into three categories of crops, weeds and soil by the manual pixel marking method, and the single classification information was placed in three different image channels, which was used for training and testing. First, a deep separable convolution neural network model based on encoder and decoder was set up. The encoder part and decoder part were merged in multi scale. The pixel location was determined by the encoder part, and the decoder part got the pixel classification. In order to solve the problem of the unbalance of the coverage rate of the classification category, the single channel standard was used. In order to control the size of the network parameters, the number of the point convolution kernel was controlled by the width multiplier and the network was used under the different resolution input conditions to control the network parameters. The model was further tested to discuss the real-time performance of the network model. Finally, we used random data enhancement technology to expand data sets, 80% of the data sets were used for training network parameters, and 20% of them were used to test network performance. 【Result】 (1) Compared with the existing pixel-wise classification method, the proposed method achieved higher classification accuracy. The average accuracy rate of the SegNet method was 90.06%, the average accuracy of the U-Net method was 92.06%, the average accuracy rate of the three channel marking training was 92.70%, and the average network accuracy of the single channel marking training was 94.99%. (2)The advantages of the single channel annotation information training method in dealing with the unbalance of classified category coverage and the less training samples were demonstrated by calculating the indexes of the single category by pixel classification by different methods. The accuracy rate of weeds pixel-wise classification SegNet method was 18.39%, U-Net method was 18.33%, the network of three channel marking training was 22.87%, and the network accuracy of single channel marking training was 41.94%. (3) The parameter size of the network model could be effectively controlled by the width multiplier. When the width multiplier was 1, the parameter size was 6.768 million, and the parameter size was reduced to 77.2 thousand when the width multiplier was 0.1. It was 1.14% of the original network parameter scale, and the accuracy rate for the classification of soil, weeds and crops was only 2.81%, 2.78% and 3.7%, respectively. According to the accuracy requirement, the scale of parameters could be further reduced. (4) Under the combined action of input resolution and width multiplier, the real-time processing capability of the network was discussed. Using GPU hardware acceleration, the rate of simultaneous recognition of three classes could reach 20 fps, and the rate of single class recognition was 60 fps. It could satisfy the real-time operation of agricultural weeding system and crop monitoring system. 【Conclusion】The pixel-wise classification method based on deep separable convolution proposed in this paper could effectively classify the soil, weeds and crops in agricultural images. At the same time, this method could deal with a single category by pixel classification in real time to meet the needs of the actual system.

Key words: crop and weed recognition, deep learning, convolutional neural networks, pixel-wise classification, semantic segmentation