Application of Efficient Channel Attention and Small-Scale Layer to YOLOv5s for Wheat Ears Detection

Abstract

Wheat is a crucial global grain crop that plays a vital role in ensuring food security worldwide. The automatic and accurate counting of wheat ears is essential for assessing wheat yield. However, the detection accuracy is greatly affected by the complex background and small target size. To address these challenges and improve the performance, we propose an enhanced YOLOv5s method. In the backbone, we introduce the efficient channel attention (ECA) to enhance the feature extraction capability of the original C3 module. Additionally, we incorporate a small-scale detection layer in the neck and prediction stages. This modification expands the original three-scale feature detection (20 × 20, 40 × 40, and 80 × 80) to a four-scale feature detection (20 × 20, 40 × 40, 80 × 80, and 160 × 160), thereby enhancing the recognition accuracy of small targets. Experimental results demonstrate that our method achieves an Accuracy (Acc) of 93.97%, which represents a 2.94% improvement over the YOLOv5s. Additionally, our method has a mean absolute error (MAE) of 0.57, a reduction of 0.6 from the YOLOv5s. The Acc of the improved YOLOv5s approaches that of YOLOv7; however, the giga floating-point operations per second (GFLOPs) and inference speed of the enhanced YOLOv5s are significantly lower than those of YOLOv7. Across various phases of the wheat test dataset, the enhanced model demonstrated superior performance. As a result, the enhanced YOLOv5s enhances its suitability for challenging field conditions and offers a dependable technical framework for ear detection and wheat yield estimation.