Domain-Adaptive Benthonic Organism Detection via Uniformizing Light Field and Color Distribution
提出DAD-ULC方案，通过编码器-解码器域转换器、水下光场感知损失和颜色分布一致性损失，统一训练与测试样本的光色域，解决水下底栖生物检测中的域偏移问题。
In this article, to exclusively conquer detection degradation of benthonic organisms due to domain shifting between training and testing scenarios, an innovative domain-adaptive detection scheme, termed DAD-ULC, is holistically invented by uniformising light field and color distribution. To that end, the encoder-decoder domain converter (EDDC) with residual connection is created, such that samples in degraded domains can be transformed into a unified domain. The underwater light field perception loss (ULFPL) is further conceptualized by virtue of a multiscale Gaussian filter, so as to directly expedite light-field conversion, getting rid of benthonic organism structure information, thereby facilitating light-domain adaptation. By exploiting the similarity between generated and referenced images in Lab space, a color distribution consistency loss (CDCL) is empowered for color-distribution transfer. Eventually, the DAD-ULC scheme is established in an end-to-end manner by integrating with EDDC, ULFPL, and CDCL modules, thereby enabling identical light-color domains between training and testing samples. Comprehensive experiments and comparisons conducted on detecting underwater objects (DUOs) and URPC2020 datasets sufficiently demonstrate effectiveness and superiority in diversified domain-shifting challenges.