Lightweight sandy vegetation object detection algorithm based on attention mechanism

Abstract

To solve the object detection task in the harsh sandy environment, this paper proposes a lightweight sandy vegetation object detection algorithm based on attention mechanism. We reduce the number of model parameters by lightweight design of the anchor-free object detection algorithm model, thereby reducing the model inference time and memory cost. Specifically, the algorithm uses a lightweight backbone network to extract features, and uses linear interpolation in the neck network to achieve multi-scale. Model algorithm compression is performed by depthwise separable convolution in the head network. At the same time, the channel attention mechanism is added to the model to further optimize the algorithm. Experiments have proved the superiority of the algorithm, the mAP in the training effect is 76%, and the prediction time per frame is 0.0277 seconds. It realizes the efficiency and accuracy of the algorithm operation in the desert environment.