Counting of Uneaten Floating Feed Pellets in Water Surface Images using ACF Detector and Sobel Edge Operator

Determination of excess feed pellet count is an essential indication of fish feeding behavior responses. To date, computer vision (CV) is considered the most practical technique to detect and count the leftover pellets. It is primarily economically viable, has broad application to other fields, and has rapid advances in computing algorithms such as deep learning (DL). This study introduces a hybrid of aggregated channel feature (ACF) detector, a non-DL object detector, and Sobel edge operator, a basic CV algorithm, to detect and count the excess floating feed pellets in water surface images with varying background noises. The ACF was used to detect the region proposal (RP) candidates of leftover pellets and discriminate the RP with low confidence scores. The selected RPs for a group of pellets, i.e., greater than 400 pixels, were further processed using the Sobel edge operator to count each pellet in the RPs. In contrast, the RPs with a pixel size of 400 are considered as a single pellet. Then, all the counted pellets in each RPs were added. This approach resulted in a considerable pellet counting estimator with r2 of 0.8 and NRMSE of 11.55% with a selected confidence score greater than 60. The main advantage of the proposed technique is that it only requires a substantially lower computational cost than a DL-based object detector.