Ocean underwater exploration is a part of oceanography that investigates the physical and biological conditions for scientific and commercial purposes. And video technology plays an important role and is extensively applied for underwater environment observation. Different from the conventional methods, video technology explores the underwater ecosystem continuously and non-invasively. However, due to the scattering and attenuation of light transport in the water, complex noise distribution and lowlight condition cause challenges for underwater video applications including object detection and recognition. In this paper, we propose a new deep encoding-decoding convolutional architecture for underwater object recognition. It uses the deep encoding-decoding network for extracting the discriminative features from the noisy low-light underwater images. To create the deconvolutional layers for classification, we apply the deconvolution kernel with a matched feature map, instead of full connection, to solve the problem of dimension disaster and low accuracy. Moreover, we introduce data augmentation and transfer learning technologies to solve the problem of data starvation. For experiments, we investigated the public datasets with our proposed method and the state-of-the-art methods. The results show that our work achieves significant accuracy. This work provides new underwater technologies applied for ocean exploration.