基于图像增强的鳕鱼识别

Image Enhancement-based Gadus Recognition

在水下目标识别的过程中,由于水下环境复杂,且水体对光线存在吸收和散射,图像获取会出现色偏、雾化以及失真等细节丢失问题,进而影响水下目标识别的精度。针对这种问题,提出一种水下图像增强的方法,即首先使用生成式对抗网络对图像进行颜色矫正和细节恢复,再利用相对全局直方图拉伸(RGHS)对图像进行对比度以及亮度调整,最大程度上还原图像本来的颜色。实验发现,采用所提算法增强后的鳕鱼图片在主观和客观评价上都得到了提升;利用添加了SE注意力机制的YOLOv5算法对增强后的鳕鱼图像进行训练识别,并与原始图像的训练结果进行比较,结果发现识别平均准确率提升了14%,训练周期数由原来的300个降低到150个,检测速度保持在143 frame/s。

Due to the complexity of underwater environment,and the absorption and scattering of water,underwater image acquisition will suffer from the loss of details such as color deviation,fog and distortion,which will affect the accuracy of underwater target recognition.To solve this problem,an underwater image enhancement method is proposed,that is,firstly,the image is color corrected and detail restored by using the generative adversarial network,and then the contrast and brightness of the image are adjusted by using RGHS,so as to restore the original color of the original image to the greatest extent.Through experiments,it is found that the gadus image enhanced by the proposed algorithm has improved in both subjective and objective evaluation.Then the enhanced gadus image is trained and recognized by the YOLOv5 algorithm with the SE attention mechanism added,and compared with the training results of the original images,it is found that the average recognition accuracy is increased by 14%,the number of training cycles is reduced from the original 300 to 150,and the detection speed remains at 143 frame/s.