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

Abstract

Recently, rapidly developing artificial intelligence and computer vision techniques have provided technical solutions to promote production efficiency and reduce labor costs in aquaculture and marine resource surveys. Traditional manual surveys are being replaced by advanced intelligent technologies. However, underwater object detection and recognition are suffering from the image distortion and degradation issues. In this work, automatic monitoring of sea cucumber in natural conditions is implemented based on a state-of-the-art object detector, YOLOv7. To depress the image distortion and degradation issues, image enhancement methods are adopted to improve the accuracy and stability of sea cucumber detection across multiple underwater scenes. Five well-known image enhancement methods are employed to improve the detection performance of sea cucumber by YOLOv7 and YOLOv5. The effectiveness of these image enhancement methods is evaluated by experiments. Non-local image dehazing (NLD) was the most effective in sea cucumber detection from multiple underwater scenes for both YOLOv7 and YOLOv5. The best average precision (AP) of sea cucumber detection was 0.940, achieved by YOLOv7 with NLD. With NLD enhancement, the APs of YOLOv7 and YOLOv5 were increased by 1.1% and 1.6%, respectively. The best AP was 2.8% higher than YOLOv5 without image enhancement. Moreover, the real-time ability of YOLOv7 was examined and its average prediction time was 4.3 ms. Experimental results demonstrated that the proposed method can be applied to marine organism surveying by underwater mobile platforms or automatic analysis of underwater videos.

Underwater Target Detection Based on Parallel High-Resolution Networks

“…With the development of computer vision and the exploitation of marine resources, biological detection in underwater environments has become a hot research topic with applications in underwater robotics [1], underwater exploration [2], and marine research [3]. However, compared to terrestrial scenes [4], underwater images suffer from problems such as color bias [5], low contrast [6], blur [7], and noise [8], resulting in the loss of clear contour and texture information in images, which limits the accuracy of target detection.…”

“…The innovations in this study are as follows: (1) aiming at the problem of complex underwater scenes and limited target feature extraction capability, a lighter high-resolution human posture estimation network (HRNet) [38] is used to enhance the target feature representation and effectively reduce the semantic information lost in the image sampling process; (2) for the problem of fuzzy and small targets, by introducing the FReLU [39] activation function, the improved convolutional block attention module (CBAM) [40] structure captures the complex feature distribution by modeling the two-dimensional space in the activation function stage to achieve the ability to model spatial information at the pixel level while performing feature enhancement in the spatial dimension and channel dimension; (3) we constructed a receptive field augmentation module (RFAB) to further obtain sufficient semantic information and rich detail information to further enhance the robustness of the features and discriminative ability of the model for underwater multi-scale target detection.…”