Object Pest Detection Method Based on Lightweight SSD_RA Algorithm.

In order to detect the types and quantities of pests in rice fields quickly and accurately, a lightweight target pest detection method SSD_RA based on SSD algorithm is proposed. In order to deal with the problems of high missed detection rate, inaccurate positioning, slow detection speed, large number of model parameters and low accuracy of the detection model, the ResNet feature extraction network was introduced and optimized. The first prediction feature layer of SSD was connected to the Conv3_x module of the ResNet network, and all network layers after the Conv3_x module were dropped. The number of parameters of the model is reduced, so that the model is more lightweight, the detection speed is improved, and the redundant features are reduced to ensure the accuracy of the model. In addition, aiming at the characteristics of small target, the structure of prediction feature layer of SSD algorithm is improved, the number of prediction feature layers is adjusted, and the output of underlying feature Conv2_x is connected to the prediction feature layer. The candidate box of each cell in the new prediction feature layer is 6, which accurately divides the boundaries of large, medium and small target boxes. The experimental results show that the mAP of SSD_RA of the improved algorithm in this paper is 84.1%, which is 23.4 percentage points higher than that of the original SSD model. The reasoning time in CPU and GPU environment is 0.056s and 0.009s, which is 0.101s and 0.005s faster than that of the original SSD model, and the model size is 51.9MB. It is reduced to about 7/100 of the original SSD model. Compared with other models, the mAP of SSD_RA is 7.8 and 4.2 percentage points higher than that of EfficientDet and RFCN, respectively. The SSD_RA model is more effective and faster to detecting insect pests and reduces the missed detection rate.