Underwater Image Enhancement Techniques for Fine Grained Visual Object Detection and Classification Using Deep Learning Architectures

Abstract

The ocean is a massive resource with military, scientific, industrial, and civil newlineuses. Close inspections of underwater images and videos may benefit many fields newlineof research, but ocean floor conditions affect the quality of underwater photography. newlineDue to poor quality, the underwater images provide certain characterstics that reduces newlinethe effectiveness of classification systems. Underwater Image Enhancement (UIE) newlinetechniques enhance the quality of underwater images. Enhancing an underwater image newlineis a difficult task due to water and lighting anomalies. Improving image processing newlinetechniques such as segmentation, object detection, and classification for UIE is the newlineprimary focus of the research. newlineUnderwater images are segmented using an interactive GrabCut algorithm. In the newlinefirst contribution of our research, a pre-processing strategy is proposed to reduce the newlinenumber of touch-ups required by the GrabCut algorithm. A contrast-limited adaptive newlinehistogram equalization (CLAHE) in the L*A*B colour space is employed after the newlinecolour correction technique to minimize the dominating colours in the image. In the newlineSingle UIE dataset (SUIE), the proposed strategy reduces the number of touch-ups by newline40%, and the algorithm s execution time is reduced by 63.68%. Though the suggested newlinemethod saves time, it still produces poor results with lower-resolution images and newlinerequires more number of manual touch-ups to achieve better results. newlineThe second contribution of the research is a fusion-based image enhancement newlinetechnique. To improve the results on low-resolution images, the image is dehazed newlineusing a pre-trained multi-layer perceptron (MLP) and then fused. Two images are newlineneeded for the fusion procedure: the contrast stretched and gamma-adjusted variants newlineof the MLP recovered image. For PCQI, UCIQE, and UIQM metrics, our method newlineoutperforms other research works on the SUIE dataset by an average score of 0.536, newline2.185, and 1.272, respectively. The YOLO object detection framework is used to newlineshow the efficiency of the proposed preprocessing technique on