Detection and Recognition of Ocean Garbage Using DIY ROV-Mounted DNN-Based Classification of Laser Images

Our work aims to contribute to a solution to difficulties in microplastic detection by investigating laser imaging for underwater garbage recognition and designing a Do-It-Yourself (DIY), low-cost underwater ROV for this purpose. The ROV is constructed using PVC pipes as its frame, which supports the electro-optics and mechanical propulsion sub-systems. A waterproof, low-cost 525 nm wavelength (green) scuba laser was strapped to the frame with eyeglass lenses used as a beam expander, while a Raspberry Pi 4 with Raspberry Pi camera was mounted on a DIY gyroscopic stabilizer and housed within a DIY waterproof box carried within the ROV to acquire laser images. Laser images were acquired under real-world conditions to evaluate the effect of lack of sunlight, distance, direct versus indirect illumination, object size and material on image quality. Several Retinex-based techniques were compared using a referenceless image quality metric to evaluate efficacy. Over 200 laser images were acquired for different distances, outdoor lighting conditions, and orientations for 7 different object classes ranging in size from microplastics to megaplastics. Two different deep learning approaches were utilized for object classification: transfer learning with VGG-16 (no laser images used in training), and a 4-layer convolutional neural network (CNN) trained exclusively on our acquired laser images. We found that the CNN trained outperformed transfer learning, yielding a 91% classification accuracy in contrast to the 75% accuracy given by VGG-16. Our results showed that a DIY underwater ROV could be used to detect and recognize garbage in oceans and lakes.