Enhanced Database Training for Small Object Detection

To effectively detect pests in large sections of orchards, this paper proposes enhancing database training to improve small object detection. The recognition rate of the proposed technique for orchard pest detection is greater than 90% in 50 x 50 x 50 cm³ blocks. Deep learning is well-suited to the variable conditions of outdoor environments. However, a large number of databases are required to strengthen the robustness of deep learning. This paper thus focuses on analyzing background complexity, the size of the detected objects, and accuracy rates to aid in the establishment of appropriate databases. The proposed database training approach effectively enhances the robustness of a detection system for pests measuring 2–5 cm at a depth of 50–70 cm in HD images with a horizontal viewing angle of 67 degrees.