Underwater Species Detection using Channel Sharpening Attention

Rethinking general underwater object detection: Datasets, challenges, and solutions

“…Generic deep detectors [25], [26], [27], [28], [29], [30] designed on high-quality images thus have poor performance when retrained on underwater images. Some researchers attempt to improve the feature representation capacity of generic deep detectors [4], [16], [17]. For example, AquaNet [4] designs two efficient components MFF and MBP using multi-scale features fusion and anti-aliasing operations.…”

“…FERNet [16] proposes a composite connection backbone and anchor refinement scheme to boost the feature representation. CSAM [17] proposes a novel channel sharpening attention module to fuse high-level image information for better feature utilization. SSoB [31] proposes a mixed antialiasing block and resorts neural architecture search to build a detection backbone for underwater scenes.…”

“…However, they suffer disappointed detection performance due to their poor feature representation in adverse conditions. Researchers thus improve the feature representation capacity of detection networks through operations such as composite connection [4], [16] or attention strategy [17]. However, this strategy does not eliminate the impact of degradation on image content such as color and texture, achieving minimal improvements.…”

In Situ Sea Cucumber Detection across Multiple Underwater Scenes Based on Convolutional Neural Networks and Image Enhancements

“…Residual networks were utilized as a backbone for enhancing the feature extraction efficiency in detecting sea cucumbers [ 19 , 20 ]. Jiang et al proposed a channel sharpening attention module (CSAM) to further fuse high-level image information [ 21 ]. The CSAM was incorporated into the YOLO v3 network and provided the network with the privilege of selecting feature maps.…”

“…Huang demonstrated that image enhancement can effectively improve the detection performance of the YOLO v5 in natural scenes [ 24 ]. A deep learning-based image restoration was applied to remove haze and light diffusion from the underwater scenes and improved the detection accuracy of sea cucumber [ 21 ]. On the other hand, synthetic images are developed as a data augmentation method to enrich the underwater image sets.…”