Forest Pest Detection Method Based on DenseNet and Pyramid Pooling

Traditional machine learning algorithms have been used in forestry to identify pests, but the results can not adapt to complex environment. In order to realize rapid and accurate pest detection on devices with limited performance, an improved lightweight model ds-YOLOv3-tiny was proposed by us. Firstly, the feature extraction network is changed to DenseNet to improve the feature extraction capability of the network and enable the model to extract high semantic features. We change the network activation function from Leaky ReLU to Swish to cover potential information loss on the network. A detection scale of 52*52 was added through tensor splicing with the second feature scale to enhance the detection ability of small objects. The k-means algorithm is used to cluster the anchors of the network, which makes the network converge more quickly. After feature extraction network, pyramid pooling module was added to improve receptive field, and target features could be learned more comprehensively. After testing, the improved model dsYOLOv3-tiny can still ensure real-time performance, and the accuracy of ds-YOLOv3-tiny is 8.3% higher than that of YOLOv3-tiny, not much different from that of YOLOv3.It is only 49.93MB in size, half the size of YOLOv3. The model dsYOLOv3-tiny in this paper combines the advantages of lightweight network and complex network to ensure the accuracy and real-time performance in forest pest detection.