# HMA-YOLO: a high precision and lightweight detection model of corn trumpet in corn precision pesticide application system

## TL;DR

This paper introduces HMA-YOLO, a lightweight and accurate model for detecting corn trumpets in corn fields to improve pesticide application precision.

## Contribution

The novel HMA-YOLO model integrates HCT, MBMS-FPN, and AMCCDH for improved detection accuracy and efficiency in corn trumpet detection.

## Key findings

- HMA-YOLO achieves a mAP@0.5 of 91.5%, precision of 89.8%, and recall of 83.7%.
- The model operates at 128 FPS with a model size of 3.1 MB and 1.407M parameters.
- HMA-YOLO outperforms mainstream detectors and is deployable on the NVIDIA Jetson Xavier NX platform.

## Abstract

Pests and diseases significantly reduce the quality and yield of corn, while the corn precision pesticide application system is one of the effective measures to solve this problem. However, the detection of corn trumpets in complex farmland environments poses significant challenges due to the high color similarity between corn trumpets and the background, the small target size, and occlusion by corn leaves.

In this paper, we propose a lightweight HMA-YOLO model to accurately detect corn trumpets in agricultural background based on YOLOv12n model. Firstly, The HCT structure that is based on CNN and Transformer architectures with assignable feature map channels is introduced into the backbone network to extract target feature information and enhance the ability of the model to distinguish between targets and backgrounds. Secondly, an efficient multi-branch and multi-scale feature pyramid network (MBMS-FPN) is developed to enhance the extraction and fusion of deep-level features of targets of varying sizes, which employs the neck heterogeneous kernel selection mechanism and feature weighted fusion module. Finally, an efficient and lightweight asymmetric multi-level channel compression detection head (AMCCDH) is improved to alleviate missed detections caused by occlusion. The AMCCDH improves detection accuracy by deepening the network path of the IoU task branch and expanding its receptive field by using 3×3 depth-wise separable convolutions. Moreover, these three improvement measures all undergo lightweight processing.

Experimental results show that HMA-YOLO achieves a mAP@0.5 of 91.5%, precision of 89.8%, and recall of 83.7%, operating at 128 FPS with only a model size of 3.1 MB and a parameter count of 1.407M. This model outperforms mainstream object detectors and has been successfully deployed on the NVIDIA Jetson Xavier NX embedded platform, which achieves real-time and efficient detection in resource-constrained environments.