Multi-scale Adaptive Learning Network with Double Connection Mechanism for Super-resolution on Agricultural Pest Images

Qingtian Zeng, Sai Chang, Shansong Wang, Weijian Ni

Abstract: Accurate recognition of pests can effectively reduce the negative impact of pests on the agricultural economy. The unprofessional shooting ways result in low-resolution images, which seriously influences the accuracy of pest recognition. The super-resolution reconstruction method can transform low-resolution images into high-resolution images. The existing super-resolution reconstruction networks have the problems such as feature missing, low feature utilization, and a large amount of computation. The Multi-scale Adaptive Learning Network (MALNet) proposed in this paper is based on residual and dense connections and consists of four modules: Rear-Compensation Enlargement Module (RCEM), Low-Frequency Feature Building Module (LFFBM), Multi-scale Details Representation Module (MsDRM), and Feature Channels Correction Module (FCCM). First, the low-frequency contour feature is built by LFFBM, and MsDRM extracts local detail features of the pest. After that, FCCM enhances important features and finally fuses with the enlarged feature map generated by RCEM. The experimental results show that, compared with other reconstruction networks, MALNet achieves an average improvement of 0.20dB and 0.011 in Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM) on Te-set and public benchmark datasets Set5, Set14, BSDS100, and Urban100. It overcomes the shortcomings of the existing networks and provides strong support for the efficient implementation of pest recognition work.