Abstract:
Recently, the potential for wheat head detection has been significantly enhanced using deep learning techniques. However, the significant challenges are variation in growth stages of wheat heads, canopy, genotype, and wheat head orientation. Furthermore, the wheat head detection task gets even more complex due to the overlapping density of wheat heads and the blur image due to the wind. For real-time wheat head detection, designing lightweight deep learning models for edge devices is also challenging. This paper proposes a lightweight WheatNet-Lite architecture to enhance the efficiency and accuracy of wheat head detection. The proposed method utilizes Mixed Depthwise Conv (MDWConv) with an inverted residual bottleneck in the backbone. Additionally, the Modified Spatial Pyramidal Polling (MSPP) effectively extracts the multi-scale features. The final wheat head bounding box prediction is achieved using WheatNet-lite Neck by utilizing Depthwise Convolution (DWConv) with a Feature Pyramid structure. It reduces 54.2 M network parameters in comparison to YOLOV3. The proposed approach outperforms the existing state-of-the-art methods with mean average precision (mAP) of 91.32 mAP@0.5 and 86.10 mAP@0.5 on GWHD and SPIKE datasets, respectively, with only 8.2 M parameters. Also, the new ACID dataset is proposed with bounding box annotation with 76.32 mAP@0.5. The experimental results are demonstrated on three different datasets viz. Global Wheat Head Detection (GWHD) , SPIKE dataset, and Annotated Crop Image Dataset (ACID) showing a significant improvement in the wheat head detection with speed and accuracy.
