基于改进SSD算法的双目鱼群图像检测

针对水下鱼群图像对比度低、鱼群尺寸不一致以及双目图像拼接出现的伪影问题,通过改进SSD (single shot MultiBox detector)算法提高图像拼接精度,实现不同尺寸鱼群快速准确检测。借助卷积层重叠相加法融合多个卷积特征,增强各个特征层的特征强度;构建特征金字塔模型,实现低卷积层的高分辨率特征与高卷积层的语义特征的融合,提高水下低对比度图像中小目标鱼群的检测精度;利用两个相同的卷积模型进行特征匹配,依据反向传播机制将第六层匹配特征逐级映射到第四层,提高特征匹配精度。在Labeled fish in the wild数据集上对本文算法进行验证,对小目标鱼群的检测精度在90%以上。

Binocular fish-school image detection based on improved SSD algorithm

Aiming at the problems such as low contrast of underwater fish-school images,inconsistent fish-school sizes and artifacts in binocular image stitching,the SSD(single shot MultiBox detector)algorithm was improved to increase the image stitching accuracy and achieve rapid and accurate detection of fish-school in different sizes.Convolution features were fused by means of the convolution layer overlapping addition method to enhance the strength of each feature.Feature pyramid was constructed to fuse the low convolutional layers with high-resolution features and the high convolutional layers with semantic features,so as to improve the detection accuracy of small targets in low-contrast underwater fish-school images.Two identical convolution models were used for feature matching.According to the backpropagation mechanism,the matching features of the sixth layer were mapped to the fourth layer step by step,and the matching accuracy of the feature was enhanced.A dataset named Labeled fish in the wild was used to verify the proposed algorithm,and the accuracy of small fish-school detection is over 90%.