As more machine learning models are now being applied in real world scenarios it has become crucial to evaluate their difficulties and biases. In this paper we present an unsupervised method for calculating a difficul...
详细信息
ISBN:
(纸本)9798350302493
As more machine learning models are now being applied in real world scenarios it has become crucial to evaluate their difficulties and biases. In this paper we present an unsupervised method for calculating a difficulty score based on the accumulated loss per epoch. Our proposed method does not require any modification to the model, neither any external supervision, and it can be easily applied to a wide range of machine learning tasks. We provide results for the tasks of image classification, image segmentation, and object detection. We compare our score against similar metrics and provide theoretical and empirical evidence of their difference. Furthermore, we show applications of our proposed score for detecting incorrect labels, and test for possible biases.
This examination intends to enhance the overall performance of welding operations through picture processing. It's going to use an aggregate of PC vision and gadgets, getting to know to perceive better and tune we...
详细信息
Robot vision servo control systems play an important role in modern automation systems, and image feature extraction and tracking, as its key components, have a direct impact on its performance and application scope. ...
详细信息
Although instance segmentation methods have improved considerably, the dominant paradigm is to rely on fully-annotated training images, which are tedious to obtain. To alleviate this reliance, and boost results, semi-...
详细信息
ISBN:
(纸本)9798350318920;9798350318937
Although instance segmentation methods have improved considerably, the dominant paradigm is to rely on fully-annotated training images, which are tedious to obtain. To alleviate this reliance, and boost results, semi-supervised approaches leverage unlabeled data as an additional training signal that limits overfitting to the labeled samples. In this context, we present novel design choices to significantly improve teacher-student distillation models. In particular, we (i) improve the distillation approach by introducing a novel "guided burn-in" stage, and (ii) evaluate different instance segmentation architectures, as well as backbone networks and pre-training strategies. Contrary to previous work which uses only supervised data for the burn-in period of the student model, we also use guidance of the teacher model to exploit unlabeled data in the burn-in period. Our improved distillation approach leads to substantial improvements over previous state-of-the-art results. For example, on the Cityscapes dataset we improve mask-AP from 23.7 to 33.9 when using labels for 10% of images, and on the COCO dataset we improve mask-AP from 18.3 to 34.1 when using labels for only 1% of the training data.
image coding for multi-task applications, catering to both human perception and machinevision, has been extensively investigated. Existing methods often rely on multiple task-specific encoder-decoder pairs, leading t...
The distinctive properties and facile integration of 2D materials hold the potential to offer promising avenues for the on-chip photonic devices, and the expeditious and nondestructive identification and localization ...
详细信息
The distinctive properties and facile integration of 2D materials hold the potential to offer promising avenues for the on-chip photonic devices, and the expeditious and nondestructive identification and localization of diverse fundamental building blocks become key prerequisites. Here, we present a methodology grounded in digital imageprocessing and deep learning, which effectively achieves the detection and precise localization of four monolayer-thick triangular single crystals of transition metal dichalcogenides with the mean average precision above 90%, and the approach demonstrates robust recognition capabilities across varied imaging conditions encompassing both white light and monochromatic light. This stands poised to serve as a potent data-driven tool enhancing the characterizing efficiency and holds the potential to expedite research initiatives and applications founded on the utilization of 2D materials. (c) 2024 Optica Publishing Group
In the past years, machine learning (ML) and deep learning (DL) have led to the advancement of several applications, including computer vision, natural language processing, and audio processing. These complex tasks re...
详细信息
ISBN:
(纸本)9798400716164
In the past years, machine learning (ML) and deep learning (DL) have led to the advancement of several applications, including computer vision, natural language processing, and audio processing. These complex tasks require large models, which is a challenge to deploy in devices with limited resources. These resource-constrained devices have limited computation power and memory. Hence, the neural networks must be optimized through network acceleration and compression techniques. This paper proposes a novel method to compress and accelerate neural networks from a small set of spatial convolution kernels. Firstly, a novel pruning algorithm is proposed based on the density-based clustering method that identifies and removes redundancy in CNNs while maintaining the accuracy and throughput tradeoff. Secondly, a novel pruning algorithm based on the grid-based clustering method is proposed to identify and remove redundancy in CNNs. The performance of the three pruning algorithms (density-based, grid-based, and partitional-based clustering algorithms) is evaluated against each other. The experiments were conducted using the deep CNN compression technique on the VGG-16 and ResNet models to achieve higher accuracy on image classification than the original model at a higher compression ratio and speedup.
The reconfiguration of machinevision systems heavily depends on the collection and availability of large datasets, rendering them inflexible and vulnerable to even minor changes in the data. This paper proposes a ref...
详细信息
ISBN:
(纸本)9798350337440
The reconfiguration of machinevision systems heavily depends on the collection and availability of large datasets, rendering them inflexible and vulnerable to even minor changes in the data. This paper proposes a refinement of Miller's Cartesian Genetic Programming methodology, aimed at generating filter pipelines for imageprocessing tasks. The approach is based on CGP-IP, but specifically adapted for imageprocessing in industrial monitoring applications. The suggested method allows for retraining of filter pipelines using small datasets;this concept of self-adaptivity renders high-precision machinevision more resilient to faulty machine settings or changes in the environment and provides compact programs. A dependency graph is introduced to rule out invalid pipeline solutions. Furthermore, we suggest to not only generate pipelines from scratch, but store and reapply previous solutions and re-adjust filter parameters. Our modifications are designed to increase the likelihood of early convergence and improvement in the fitness indicators. This form of self-adaptivity allows for a more resource-efficient configuration of image filter pipelines with small datasets.
The automated detection and classification of plant diseases based on images of leaves is a significant milestone in agriculture. Due to the increasing popularity of digital imageprocessing, machine learning, and com...
详细信息
The automated detection and classification of plant diseases based on images of leaves is a significant milestone in agriculture. Due to the increasing popularity of digital imageprocessing, machine learning, and computer vision techniques, it has been proposed that these could be used for the early detection of diseases. However, the accuracy of these techniques is still considered to be a challenge. In this paper, the concept of deep learning was used to identify and predict cotton plant disease status using images of leaves and plants collected in an uncontrolled environment. This paper focuses on solving the problem of cotton plants disease detection and classification using an improved Deep Convolution Neural Network based model. Three different experimental configurations were investigated to study the impacts of different data split ratios, different choices of pooling layer (max-pooling vs. average-pooling), and epoch sizes. The models were trained using a database of 2293 images of cotton leaves and plants. The data included four distinct classes of leaves, plant disease combinations, and their respective categories. For classifying leaves and plant diseases in cotton plants, our model attained an accuracy of 97.98%. The proposed technique outperformed the recent approaches indicated in earlier literature for relevant parameters. As a result, the technique is intended to reduce the time spent identifying cotton leaf disease in significant production regions and human error and the time spent determining its severity.
Geological hazards in densely vegetated mountainous regions are challenging to detect due to terrain concealment and the limitations of traditional visualization methods. This study introduces the LiDAR image highligh...
详细信息
Geological hazards in densely vegetated mountainous regions are challenging to detect due to terrain concealment and the limitations of traditional visualization methods. This study introduces the LiDAR image highlighting algorithm (LIHA), a novel approach for enhancing micro-topographical features in digital elevation models (DEMs) derived from airborne LiDAR data. By analogizing terrain profiles to non-stationary spectral signals, LIHA applies locally estimated scatterplot smoothing (Loess smoothing), wavelet decomposition, and high-frequency component amplification to emphasize subtle features such as landslide boundaries, cracks, and gullies. The algorithm was validated using the Mengu landslide case study, where edge detection analysis revealed a 20-fold increase in identified micro-topographical features (from 1907 to 37,452) after enhancement. Quantitative evaluation demonstrated LIHA's effectiveness in improving both human interpretation and automated detection accuracy. The results highlight LIHA's potential to advance early geological hazard identification and mitigation, particularly when integrated with machine learning for future applications. This work bridges signal processing and geospatial analysis, offering a reproducible framework for high-precision terrain feature extraction in complex environments.
暂无评论