Intelligent feeding control of fish shoal using image texture features and decision tree

Aquaculture feeding has been required to impelment sustainable practices in recent years. Manual feeding cannot fully meet the requirement of large-scale production, due to the high labor intensity and low feeding precision. In this study, an intelligent feeding control system was proposed using image texture analysis. A decision tree model was established for the classification of low or high feeding intensity levels. Gabor filters were integrated as the first step to enhance the image texture. Gray-level co-occurrence matrix (GLCM), gray-level difference squared statistics (GLDS), and histogram statistical analysis (Histogram) were utilized to capture the multi-dimensional texture features which can reflect feeding intensity levels. A decision tree model was then constructed using the extracted texture feature dataset. A series of experiments were conducted on the Pelteobagrus fulvidraco in a recirculating aquaculture system from the Huazhong Agriculture University in Wuhan City of China. In the experiment, the video data was captured 2 min before feeding and 2 min after feeding in each feeding round to serve as a training dataset. The video frames in the traning dataset were then processed using Gabor filters, in order to enhance the texture features, particularly those related to fish feeding intensity levels. The Gabor filter was used to capture the local texture information and then suppress noise. The scale and direction were also optimized to maximize the differentiation between feeding and non-feeding states. The enhanced images were then subjected to texture feature extraction using GLCM, GLDS, and Histogram. The statistical matrics of those texture patterns, such as the contrast, and entropy relevant features, were computaed to feed a decision tree model to discriminate the feeding intensity levels. The decision tree model was trained to classify the fish-feeding states into two categories: high-density feeding (H) and low-density non-feeding (L), according to the texture feature extracted. The cross-validation was applied to evaluate the accuracy of the proposed approach. The results demonstrated that the classification accuracy reached 98.96% in the training set and 95.83% in the testing set. The real-time feeding control was further realized using the decision tree model immediately after the model construction was finished. The whole cycle of a feeding process was divided into three stages: pre-feeding, initial feeding, and intelligent control. During the pre-feeding stage, the video data was collected and the image frames were labeled as being in non-feeding state. In the initial feeding stage, the feeding state was labeled as being in feeding state. Both datasets collected in the pre-feeding and initial feeding stage were used to extract the image texture featuers and construct the decision tree model. In the intelligent control stage, the build decision tree model was applied to determine when to stop feeding, according to the texture features collected by the real-time vedio data. The control system dynamically adjusted the feeding process. As such, the feed was delivered only when the fish was actively feeding, thus preventing the overfeeding or underfeeding. The computational efficiency of the control system was evaluated on the entire process, including Gabor filtering, texture feature extraction, and decision tree construction. The deployment was tested on a standard laptop with an Intel Core i7-8550U processor and 16 GB of RAM. The video frames were processed in real-time using the control system and each image frame was processed in terms of the texture feature computation was less than 0.1 seconds. The high efficiency of the system was suitable for practical implementation in aquaculture settings. Several advantages were also gained over the existing approaches. Firstly, the Gabor filters and texture features provided high robustness against the variations in the lighting and water conditions, which were the common challenges in aquaculture environments. Secondly, the decision tree model offered the interpretability to understand the decision-making process for the parameters adjustment as needed. Thirdly, the real-time training and prediction strategy can be applied by different aquaculture scenarios and species. Compared with the convolutional neural networks and optical flow techniques, comparable or superior accuracy was achieved with the less intensive and more interpretable computation. In conclusion, the intelligent system of feeding control was a viable solution to improve feed efficiency with labor cost-saving in aquaculture, according to image texture analysis and decision tree modeling. A promising tool was obtained to dynamically adapt under different feeding scenarios. The high accuracy was also achieved to detect the feeding behavior.