Deep recursive super resolution network with Laplacian Pyramid for better agricultural pest surveillance and detection

Abstract Computer vision technologies greatly improved the efficiency of recognizing and controlling of agricultural pests. However, the density of cameras deployed in the farmland is usually sparse and the images or videos of agricultural pests collected are often obscure. This always results in low resolution of pests in the pictures, making them difficult to observe and monitor. In addition, the existing object detection method is not satisfactory for the detection of small targets with low pixel resolution. Therefore, it is necessary to restore and upsample the collected images so as to improve the recall rate of the detection. In this work, we proposed a novel super resolution model based on deep recursive residual network. Compared with the traditional interpolation methods and the models with shallow convolutional neural networks, the method we proposed is more powerful in image reconstruction and achieves the state of the art performance. The experimental results show that our method greatly improved the recall rate of pest detection by 202.06%. In addition, compared with image upscaling methods such as Bicubic Interpolation and Super-Resolution Convolutional Neural Network (SRCNN), our method is average 111.31% and 41.89% improved respectively. The model we put forward could reduce the density of the camera layout of the agricultural Internet of Things (IOT) monitoring systems and reduce cost of infrastructure, which is of high practical value.