采用改进SSD网络的海参目标检测算法

随着海参养殖业快速发展，利用水下机器人代替人工作业的海参智能捕捞已成为发展趋势。浅海环境复杂，海参体色与环境区分性差、海参呈现半遮蔽状态等原因，导致目标识别准确率低下。此外由于景深运动，远端海参作为小目标常常未被识别成功。为解决上述问题，该研究提出一种基于改进SSD网络的海参目标检测算法。首先通过RFB（Receptive Field Block）模块扩大浅层特征感受野，利用膨胀卷积对特征图进行下采样，增加海参细节、位置等信息，并结合注意力机制，对不同深度特征进行强化，将计算得出的权重与原特征信息相乘以此获得特征图，使结果包含最具代表性的特征，也抑制无关特征。最后实现特征图融合，进一步提升水下海参的识别精度。以实际拍摄的视频进行测试验证，在网络结构层面上，对传统算法进行改进。试验结果表明，基于改进的SSD网络的海参目标检测算法的平均精度均值为95.63%，检测帧速为10.70帧/s，相较于传统的SSD算法，在平均精度均值提高3.85个百分点的同时检测帧速仅减少2.8帧/s。与Faster R-CNN算法和YOLOv4算法进行对比试验，该研究算法在平均精度均值指标上，分别比YOLOv4、Faster R-CNN算法提高4.19个百分点、1.74个百分点。在检测速度方面，该研究算法较YOLOv4、Faster R-CNN算法分别低4.6帧/s、高3.95帧/s，试验结果表明，综合考虑准确率与运行速度，改进后的SSD算法较适合进行海参智能捕捞任务。研究结果为海参智能捕捞提供参考。

Algorithm for detecting sea cucumbers based on improved SSD

Abstract: Intelligent fishing of sea cucumbers has been an ever-increasing trend using underwater robots in recent aquaculture, instead of the conventional manual operations. However, there is a relatively low distinction between the sea cucumbers and the complex living environment, some of which are semi-hidden in the natural ocean. It is easy to induce the low accuracy of an underwater robot in the recognition of the sea cucumber targets. Particularly, the remote sea cucumbers cannot be recognized as the small targets with the depth of field during the movement of an underwater robot in the natural environment. In this study, the object detection algorithm was proposed for the sea cucumbers using improved Single Shot multibox Detector (SSD) network deep learning. Firstly, the shallow-feature receptive field was improved to increase the location information using a receptive field block. The spatial attention and channel attention mechanisms were then combined to strengthen the features of different depths in the network. The original feature information was multiplied to obtain the weight between each feature channel and feature space. As such, the most representative features were achieved in the channel and spatial feature maps without the irrelevant features. Finally, the fusion of the feature map was performed to further improve the precision of sea cucumber recognition. The actual video was taken to verify the model during testing in the experiment. The improved recognition rate of underwater sea cucumber was obtained at the level of network structure. The experimental results show that the Mean Average Precision (mAP50) was 95.63% for the target detection of sea cucumber using the improved SSD network, and the detection frame rate was 10.7 frame/s. Specifically, the mAP50 increased by 3.85 percentage points, while the detection frame rate was only reduced by 2.8 frame/s, compared with the traditional SSD. The precision-recall (P-R) curves were compared before and after the model improvement. There was a larger area between the P-R curve of the improved SSD model and the X and Y coordinate axes, and the balance point was closer to the coordinates (1, 1), indicating the better performance of the improved SSD model. The Faster R-CNN and YOLOv4 were selected to verify the effectiveness of the improved SSD. The mAP50 values of the improved model were 4.19 and 1.74 percentage points higher than those of the YOLOv4 and Faster R-CNN, respectively, indicating the better system performance of the improved model on the P-R curve than those algorithms. The detection speed of the improved model was 4.6 frame/s lower than that of YOLOv4, whereas, that was 3.95 frame/s higher than that of Faster R-CNN. Consequently, the improved SSD was more suitable for the underwater robot of sea cucumber in the intelligent fishing task, considering the target detection accuracy and running speed. The finding can provide a strong reference for the intelligent fishing of sea cucumbers in aquaculture.