

The objective of canal semantic segmentation in this work preliminarily investigates the visual intelligence on an SSSB (Self-Sailing Sweeper Boat) that can visualize constraint sailing in a restricted region further for sweeping the canal. The contribution and novelty of this work firstly proposes the night-scenario canal segmentation in terms of the low-lightness problems, in which the boundaries between the canal region and the bank shore cannot be precisely distinguished for ground-true labeling prior to the dataset training and validation, further decreasing the segmentation efficiency. To do so, this work investigates the contrast enhancement approach on both Histogram Equalization (HE) and Adaptive Histogram Equalization (AHE) methods, respectively, naming AH2E2 on night-scenario canal to increase boundary visibility to benefit the precise labeling of the ground-true region for canal segmentation. Thereafter, three U-net-based approaches, including the Primordial U-Net, ResU-Net, and AttresU-net, of which six combinations from HE and AHE are examined for evaluation. To further inspect the tradeoff between the training cost and the segmentation efficiency in terms of required EPOCH and the engaged ground-true number. The experiments dynamically set the participated ground-true labels from 150 to 750 step 150 and set EPOCH as 100. Experimental results reported that HE for contrast enhancement method with AttresU-net learning approach performed superior segmentation efficiency compared to the other five combinations. However, HE+AttresU-net herein is observed to give a higher training cost than the other five combinations. More discussions are elucidated in the experiments.