Image Analysis based on AI application for Manufacturing Automation

Abstract

Abalone detection and counting systems via conveyor belts using machine learning have become a critical and most vital aspect in the fishing industry. In our paper, we present a detection and counting approach for Multi-Object Tracking (MOT) from abalone video data. Considering computational effectiveness and improved detection algorithms, Multiple Object and Tracking is one of the most sort out paradigms. You only look once (YOLOV3) with Tensor Flow algorithm built based on Deep SORT algorithm mainly implemented for MOT specification. Other methods used had low effectiveness with high operational time and wrong counting rate because some of the abalones stick to each other most of the time. However, to elaborate the challenges posted by the video data of abalone entangling and pairing together on multiple occasions which is hard to count, we tested with the open source-based algorithm (OpenCV) which did detection but incorrect tracking and sorting. After combining Yolov3 and Tensor Flow for generating detections in each frame based on the model (Darknet-53), we use depth separable convolutions and pointwise group convolutions to reduce the parameter size of the network, the results of our experiment exhibited competitive performance in comparison with the open source-based library OpenCV. This proposed algorithm can apply in various conveyors counting controlled-based systems. We also propose a technique for classifying normal and abnormal images from a broad class of Convolutional Neural Network (CNN) models, data classification more clear and explainable. We introduced Gradient-weighted Class Activation Mapping (Grad-CAM) that uses the gradients of any objective approach of normal or abnormal images fundamentally into the ultimate convolutional layer to yield a coarse localization map for highlighting the critical sectors in the image for predicting the approach. In contrast to previous approaches, Grad-CAM applies to a wide variety of CNN models: CNNs with fully connected layers (e.g. ResNet50 model) and CNN models used in line with various inputs (e.g. visualization) or reinforcement learning, without architectural transformation or re-training. We incorporate Grad-CAM with current existing fine grained visualizations to generate a high-resolution class discriminative factor for visualization. Guided Grad-CAM is then applied for image classification, captioning based on pre-trained ResNet50 architectures. For image captioning, our visualizations show that even ResNet50 based models learn to localize discriminative regions of the input image. We develop identification of essential neurons via Grad-CAM and incorporate it with neuron names to provide textual explanations for model decisions. Finally, the used test data to establish appropriate trust in predictions from the model and show that Grad-CAM can give accurate predictions from deep networks and show that Grad-CAM helps to make a more accurate prediction even close to identical images either normal or abnormal.