Rapid and Accurate Identification of Grass Seedlings in Agricultural Fields Based on Optimized YOLOX Model

Abstract

Traditional agricultural cultivation is labor-intensive and vulnerable to natural climate conditions, such as heavy rainfall and drought. Concerns over food safety have also brought attention to the growth of weeds and the misuse of agricultural chemicals, which can have a serious negative impact on crop growth and safety. We investigated the feasibility of the YoloX model in the field of agricultural weed identification to address the problem of weed handling in growing crops. In order to overcome the effects of climate and environment, we chose a purchased weed model for our study. We used a binocular vision camera for image acquisition and created a database containing 6,000 samples and enhanced the original database of 1,000 samples with data. In order to address the complex background of the weed images, the changing lighting environment of the binocular camera-acquired images, and the noise interference, we performed histogram equalization, image denoising, and background processing on the dataset. These processing measures aim to improve the overall learning efficiency of the model in order to improve the accuracy of deep learning on weeds. Target detection platforms based on the TensorFlow and PyTorch frameworks were established, respectively, and the mainstream target detection models Faster R-CNN and YoloX series target detection models were trained with the same dataset for comparative analysis using the longitudinal comparison method. The results show that the training under the PyTorch framework yields better models than the training under the TensorFlow framework. YoloX-x has higher recognition accuracy, faster recognition speed, and more stable compared to the Faster R-CNN model, with an average recognition rate of 97.07% and an average recognition time of 0.062 s. Moreover, the optimization of YoloX by incorporating an attention-focusing mechanism resulted in an improved accuracy rate with a decrease in recognition time, with an average recognition rate of 97.70% and an average recognition time of 0.029 s.