Seashore Debris Detection Model with KaKaXi Camera Custom Dataset Using Instance Segmentation

Open Access

Abstract: Marine debris is impacting coastal landscapes majorly by affecting biodiversity, impairing recreational uses, causing losses to fishing industries, maritime industries, etc. Motivated by the need for automatic and cost-effective approaches for debris monitoring and removal, we employed computer vision technique together with deep learning-based model to identify and classify marine debris on several beach locations. This paper provides a comparative analysis of state-of-the-art deep learning architectures and proposed architecture which is used as feature extractor for debris image classification.

The model is being proposed to detect seven categories of marine debris using a custom debris dataset, with the help of instance segmentation and a shape matching network, which can then be cleaned timely and efficiently. The manually constructed dataset for this system is created by annotating fixed KaKaXi camera images using CVAT with seven types of labels. A pre-trained HOG shape feature extractor is being used on LIBSVM along with template matching to improve the predicted masked images obtained via Mask R-CNN training. This system intends to timely alert the cleanup organizations with the recorded live debris data. The proposed network resulted in the improvement of misclassification of debris masks for objects with different illuminations, shape, occlusion and viewpoints.

Keywords: debris; fixed camera images; computer vision; instance segmentation; deep learning; template matching; Histogram of Gradients (HOG)