An Improved Efficient Convolutional Neural Network for Weed Seedlings Detection

Abstract

The growth of weeds in the fields is one of the important factors affecting crop yields. Timely detection and controlling of weeds have a great positive effect on the healthy growth of crop seedlings. Identifying the types of weeds correctly can effectively improve the efficiency of weed removal. Convolutional neural network is a good method for the detection of weed seedlings. With a suitable convolutional neural network model, the types of seedlings can be classified through pictures, which improves the efficiency of agricultural work greatly. This paper constructs a convolutional neural network model based on MobileNet and TensorFlow. The model is trained by inputting pictures of crops and weeds seedlings in 12 different types. The training model is evaluated with performance of 96.88% in identifying seedling types. Due to the features of high efficiency and lightweight of MobileNet, this model can be better applied to mobile devices than others, which is convenient for agricultural workers to use.