Underwater sea cucumbers detection based on pruned SSD

Abstract

We proposed method for sea cucumber detection based on SSD (Single Shot Multibox Detector) and convolution kernel pruning. Sea cucumbers images taken by underwater robot were enhanced by multi-scale Retinex firstly, which can make images clearer and enhance the details. A dataset was built based on data augmentation, there were 5530 labelled original images in the dataset totally and SSD was trained to detect the sea cucumbers. The redundant filters of the feature extraction network VGG16 were pruned based on convolution kernel pruning. The parameter quantity and calculation of the model were effectively reduced without loss of precision. The mean average precision of pruned SSD reached 89.54% on test dataset. The detection precision for underwater sea cucumbers was 94. 32% and the recall ratio was 91.09%. Compared with original model, the detection speed was increased by 20.52% to 46.4 frame/s. Compared with the detection performance of the pruned SSD in different object scale, different weather conditions, and other methods for object detection, the pruned SSD can reach a high level of both detection precision and speed, which also has better robustness. The pruned SSD proposed in this paper can realize the accurate and real time detection of sea cucumbers.