Underwater Sea Cucumbers Detection Based on Improved SSD

Abstract

A sea cucumbers detection method based on SSD and depthwise separable convolution is proposed. Firstly, sea cucumbers images taken by underwater robot are enhanced by Multi Scale Retinex, which can make underwater images clearer. MobileNet-SSD which uses a 13-layer depthwise separable convolution as basic feature extractor is trained to detect the sea cucumbers and model quantification is applied to the model to further reduce the model size and improve detection speed. Experimental results demonstrate that the mean average precision of improved SSD reaches 89.41 %, the detection precision for underwater sea cucumbers is 93.76% and the recall ratio is 91.17%, the detection speed can reach 19.8f/s in CPU mode which is no longer limited by the performance of computer. Compared with other methods for object detection, the improved SSD can reach a high level of both detection precision and speed, which can realize the rapid and accurate detection of sea cucumbers.