Research on the identification and distribution of biofouling using underwater cleaning robot based on deep learning

Abstract

The rapid identification of biofouling is of great significance to the intelligent operation and maintenance of large ships or other deep-sea and offshore projects. The biofouling's optical imaging properties and its dispersion and concealment bring challenges to the correct segmentation of its images. In this article, a filter-guided inverse dark channel inversion exposure compensation (FIDCE) algorithm is proposed for the ambiguous image by exploring an underwater optical imaging model partially over-exposed and partially under-exposed. Also, MFONet, a pixel-level segmentation model, is proposed. Its backbone feature extraction network includes MobileNetV2—a lightweight CNN model with the Sandglass block as the basic unit, and combines with ASPP enhanced feature extraction module, thereby enhancing the network perception field. In this model, SENet is also used to enhance the deep and shallow feature fusion. In addition, the underwater image evaluation functions are compared to verify the validity of the image processing algorithm mentioned above. According to the results, a more accurate and quicker effect can be observed in the fine-grained identification of biofouling using MFONet. Nowadays, such an algorithm can be used for a priori study of biofouling and the cleaning operation of the robot.