# Research on Field Weed Target Detection Algorithm Based on Deep Learning

## TL;DR

This paper introduces SSS-YOLO, a new deep learning algorithm for detecting field weeds, especially in cases where weeds are occluded or overlapping, to improve smart agriculture.

## Contribution

The paper proposes SSS-YOLO, a novel algorithm with modules to enhance detection accuracy for occluded or overlapping weeds.

## Key findings

- SSS-YOLO outperforms existing algorithms in detecting occluded or overlapping weeds.
- The SCB, SPPF EGAS, and EMSN modules improve feature extraction and contextual reasoning.
- Experiments on self-built and public datasets show significant performance improvements.

## Abstract

Weed detection algorithms based on deep learning are considered crucial for smart agriculture, with the YOLO series algorithms being widely adopted due to their efficiency. However, existing YOLO algorithms struggle to maintain high accuracy, while low parameter requirements and computational efficiency are achieved when weeds with occlusion or overlap are detected. To address this challenge, a target detection algorithm called SSS-YOLO based on YOLOv9t is proposed in this paper. First, the SCB (Spatial Channel Conv Block) module is introduced, in which large kernel convolution is employed to capture long-range dependencies, occluded weed regions are bypassed by being associated with unobstructed areas, and features of unobstructed regions are enhanced through inter-channel relationships. Second, the SPPF EGAS (Spatial Pyramid Pooling Fast Edge Gaussian Aggregation Super) module is proposed, where multi-scale max pooling is utilized to extract hierarchical contextual features, large receptive fields are leveraged to acquire background information around occluded objects, and features of weed regions obscured by crops are inferred. Finally, the EMSN (Efficient Multi-Scale Spatial-Feedforward Network) module is developed, through which semantic information of occluded regions is reconstructed by contextual reasoning and background vegetation interference is effectively suppressed while visible regional details are preserved. To validate the performance of this method, experiments are conducted on both our self-built dataset and the publicly available Cotton WeedDet12 dataset. The results demonstrate that compared to existing algorithms, significant performance improvements are achieved by the proposed method.