A novel end-to-end digital image processing pipeline for intelligent visual perception and decision-making systems

In digital image-processing systems, typically the different phases of image pre-processing, feature extraction, visual recognition and decision-making are performed in a sequential manner, thereby introducing data redundancy and restricting the ability to provide actionable outputs. This study aims to propose an end-to-end digitally integrated image processing pipeline for acquiring, noise reducing and contrast enhancing images, performing hybrid feature extraction, visual recognition, confidence evaluation and decision generation as a unified pipeline. The proposed feature-extraction module partly makes use of hand-crafted edge descriptions and texture ones together with hierarchical deep features, and the decision module transforms the recognition results to actions, alerts, and further analysis requests according to prediction confidence and context rules. It was benchmarked on CIFAR-10, MNIST, an ImageNet subset, and on a chest X-ray dataset, against ResNet50, EfficientNet-B0, Vision Transformer and MobileNetV3. The proposed pipeline achieved accuracy as 0.94, 0.99, 0.91 and 0.93 for four data sets, respectively, with a precision of 0.93, recall of 0.92, F1 Score of 0.925 and AUC of 0.97. It also needed about 13 ms per image for a model size of 38 Mb, which marked a pleasant trade-off between prediction performance and computation costs. The ablation analysis, on the other hand, revealed that each of the factors of preprocessing, feature fusion, and confidence-aware decision generation had an impact on the overall functioning of the system. Sensitivity analysis additionally demonstrated that each of the factors: preprocessing, feature fusion, and confidence-aware decision generation has an impact on the functioning of the overall system. The results show the capability of this proposed system in fast visual perception and decision support application in industrial inspection, surveillance, robotics and medical-image analysis. The present assessment is based upon publicly available benchmark datasets, but it is necessary to investigate it separately in real world application.