# LPD-YOLOv7-tiny: An enhanced lightweight YOLOv7-tiny model for real-time potato quality detection

## TL;DR

This paper introduces LPD-YOLOv7-tiny, a lightweight model for real-time potato quality detection that improves accuracy and speed.

## Contribution

The novel integration of MobileNetV3, BiFormer, SimAM, and Focal-EIOU loss in a YOLOv7-tiny framework for potato quality detection.

## Key findings

- LPD-YOLOv7-tiny achieves 90.3% mAP with 5.8 MB model size and 142.5 fps inference speed.
- The model outperforms mainstream detection models in accuracy and computational efficiency.
- It effectively handles complex backgrounds in potato quality detection tasks.

## Abstract

To solve the problems of low detection accuracy, large model size and slow reasoning speed of existing potato quality detection models, this paper proposes LPD-YOLOv7-Tiny, a lightweight potato sprout and spoilage detection model based on YOLOv7-Tiny. The proposed model introduces MobileNetV3 small, BiFormer, SimAM, and the Focal-EIOU loss function. MobileNetV3 small greatly reduces the number of parameters and computational complexity of the model, BiFormer enhances the multi-scale feature fusion capability of the model, and the SimAM module effectively suppresses irrelevant information and strengthens local features. The Focal-EIOU loss function improves the model’s attention to difficult classification samples and enhances its bounding box regression capability. LPD-YOLOv7-Tiny achieves excellent detection performance on potatoes under complex background conditions: mAP is increased to 90.3%, the number of parameters is reduced to 5.8 MB, the number of computations is reduced to 10.1 G, and the inference speed is increased to 142.5 fps. Compared with mainstream detection models such as the YOLO Basic series, SSD and speed-RCNN, LPD-YOLOv7-Tiny achieves significantly improved performance in terms of detection accuracy, positioning capability and computational efficiency, indicating it has wide application potential in resource-constrained and high-precision scenarios.