Abstract Image analysis technology has drawn dramatic attention and developed rapidly because it enables a non-extractive and non-destructive approach to data acquisition of crab aquaculture. Owing to the irregular shape, multi-scale posture and special underwater environment, it is very challenging to adopt the traditional image recognition methods to detect crabs quickly and effectively. Consequently, we propose a real-time and robust object detector, Faster MSSDLite, for detecting underwater live crabs. Lightweight MobileNetV2 is selected as the backbone of a single shot multi-box detector (SSD), and standard convolution is replaced by depthwise separable convolution in the prediction layers. A feature pyramid network (FPN) is adopted at low extra cost to improve the detection precision of multi-scale crabs and make up for the deficiency of SSD to force different network layers to learn the same features. More significantly, the unified quantized convolutional neural network (Quantized-CNN) framework is applied to quantify the error correction of the improved detector for further accelerating the computation of convolutional layers and compressing the parameters of fully-connected layers. The test results show that Faster MSSDLite has better performance than traditional SSD. The average precision (AP) and F1 score of detection are 99.01% and 98.94%, respectively. The detection speed can reach 74.07 frames per second in commonly configured microcomputers (~8× faster than SSD). The computation amount of floating-point numbers required by the detection is reduced to only 0.32 billion (~49× smaller than SSD), and the size of the model is compressed into 4.84 MB (~28× smaller than SSD). The model is also more robust, which can stably detect underwater live crabs in real-time, estimate the live crab biomass in water bodies automatically, and provide reliable feedback information for the fine feeding of automatic feeding boats.

中文翻译：

---

基于深度学习的水下活蟹实时鲁棒检测器

摘要 图像分析技术由于能够以非提取、无损的方式采集蟹类养殖数据而受到广泛关注并得到迅速发展。由于形状不规则、多尺度姿态和特殊的水下环境，采用传统的图像识别方法快速有效地检测螃蟹是非常具有挑战性的。因此，我们提出了一种实时且强大的物体检测器 Faster MSSDLite，用于检测水下活蟹。轻量级 MobileNetV2 被选为单镜头多盒检测器 (SSD) 的主干，在预测层中标准卷积被深度可分离卷积取代。采用低额外成本的特征金字塔网络（FPN）来提高多尺度螃蟹的检测精度，弥补SSD强制不同网络层学习相同特征的不足。更重要的是，统一量化卷积神经网络（Quantized-CNN）框架被应用于量化改进检测器的纠错，以进一步加速卷积层的计算和压缩全连接层的参数。测试结果表明，Faster MSSDLite 的性能优于传统SSD。检测的平均精度 (AP) 和 F1 分数分别为 99.01% 和 98.94%。在常用配置的微型计算机中，检测速度可以达到每秒 74.07 帧（比 SSD 快 8 倍）。检测所需的浮点数计算量减少到只有 3.2 亿（比 SSD 小 49 倍），模型大小压缩到 4.84 MB（比 SSD 小 28 倍）。该模型也更加稳健，能够稳定实时检测水下活蟹，自动估算水体中活蟹生物量，为自动投喂船的精细投喂提供可靠的反馈信息。