Lightweight improvement study of sonar image detection network
Yan Zhou, Shaochang Chen · 2021
For the task of underwater sonar target detection, a lightweight convolutional neural network design approach for detection tasks is investigated in order to enable the deployment of detection networks to edge platforms. MobileNet-SSD, a single-stage detection framework with MobileNet as the backbone network, was selected as the benchmark network on which to continue optimisation and improvement. The RFB module containing the dilated convolution is introduced for fitness improvement and the RFB_i module is proposed. Using the RFB_i module to optimise the MobileNet-SSD, IMSRNet is proposed. The experimental results show that IMSRNet's detection precision on the SCTD dataset is 2.84% higher than MobileNet-SSD, its inference speed is 11.34% faster than MobileNet-SSD, the number of parameters is 84.96% of that of MobileNet-SSD, and the computation is 82.47% of that of MobileNet-SSD. The proposed algorithm was experimentally verified to outperform MobileNet-SSD on the SCTD dataset.