Convolutional neural network-based automatic image recognition for agricultural machinery

An internet of things-based subsoiling operation monitoring system for agricultural machinery is able to identify the type and operating state of a certain machinery by collecting and recognizing its images; however, it does not meet regulatory requirements due to a large image data volume, heavy workload by artificial selective examination, and low efficiency. In this study, a dataset containing machinery images of over 100 machines was established, which including subsoilers, rotary cultivators, reversible plows, subsoiling and soil-preparation machines, seeders, and non-machinery images. The images were annotated in tensorflow, a deep learning platform from Google. Then, a convolutional neural network (CNN) was designed for targeting actual regulatory demands and image characteristics, which was optimized by reducing overfitting and improving training efficiency. Model training results showed that the recognition rate of this machinery recognition network to the demonstration dataset reached 98.5%. In comparison, the recognition rates of LeNet and AlexNet under the same conditions were 81% and 98.8%, respectively. In terms of model recognition efficiency, it took AlexNet 60 h to complete training and 0.3 s to recognize 1 image, whereas the proposed machinery recognition network took only half that time to complete training and 0.1 s to recognize 1 image. To further verify the practicability of this model, 6 types of images, with 200 images in each type, were randomly selected and used for testing; results indicated that the average recognition recall rate of various types of machinery images was 98.8%. In addition, the model was robust to illumination, environmental changes, and small-area occlusion, and thus was competent for intelligent image recognition of subsoiling operation monitoring systems.