Smooth Group L1/2 Regularization for Pruning Convolutional Neural Networks

In this paper, a novel smooth group L1/2 (SGL1/2) regularization method is proposed for pruning hidden nodes of the fully connected layer in convolution neural networks. Usually, the selection of nodes and weights is based on experience, and the convolution filter is symmetric in the convolution neural network. The main contribution of SGL1/2 is to try to approximate the weights to 0 at the group level. Therefore, we will be able to prune the hidden node if the corresponding weights are all close to 0. Furthermore, the feasibility analysis of this new method is carried out under some reasonable assumptions due to the smooth function. The numerical results demonstrate the superiority of the SGL1/2 method with respect to sparsity, without damaging the classification performance.

EFFICIENT LINE DETECTION METHOD BASED ON 2D CONVOLUTION FILTER

The article proposes an efficient line detection method using a 2D convolution filter. The proposed method was compared with the Hough transform, the most popular method of straight lines detection. The developed method is suitable for local detection of straight lines with a slope from -45˚ to 45˚. Also, it can be used for curve detection which shape is approximated with the short straight sections. The new method is characterized by a constant computational cost regardless of the number of set pixels. The convolution is performed using the logical conjunction and sum operations. Moreover, design of the developed filter and the method of filtration allows for parallelization. Due to constant computation cost, the new method is suitable for implementation in the hardware structure of real-time image processing systems.

Retinal Blood Vessel Extraction Using a New Enhancement Technique of Modified Convolution Filters and Sauvola Thresholding

The retinal blood vessels in humans are major components with different shapes and sizes. The extraction of the blood vessels from the retina is an important step to identify the type or nature of the pattern of the diseases in the retina. Furthermore, the retinal blood vessel was also used for diagnosis, detection, and classification. The most recent solution in this topic is to enable retinal image improvement or enhancement by a convolution filter and Sauvola threshold. In image enhancement, gamma correction is applied before filtering the retinal fundus. After that, the image should be transformed to a gray channel to enhance pictorial clarity using contrast-limited histogram equalization. For filter, this paper combines two convolution filters, namely sharpen and smooth filters. The Sauvola threshold, the morphology, and the medium filter are applied to extract blood vessels from the retinal image. This paper uses DRIVE and STARE datasets. The accuracies of the proposed method are 95.37% for DRIVE with a runtime of 1.77[Formula: see text]s and 95.17% for STARE with 2.05[Formula: see text]s runtime. Based on the result, it concludes that the proposed method is good enough to achieve average calculation parameters of a low time quality, quick, and significant.

A Semi-Automatic Workflow to Extract Irregularly Aligned Plots and Sub-Plots: A Case Study on Lentil Breeding Populations

Plant breeding experiments typically contain a large number of plots, and obtaining phenotypic data is an integral part of most studies. Image-based plot-level measurements may not always produce adequate precision and will require sub-plot measurements. To perform image analysis on individual sub-plots, they must be segmented from plots, other sub-plots, and surrounding soil or vegetation. This study aims to introduce a semi-automatic workflow to segment irregularly aligned plots and sub-plots in breeding populations. Imagery from a replicated lentil diversity panel phenotyping experiment with 324 populations was used for this study. Image-based techniques using a convolution filter on an excess green index (ExG) were used to enhance and highlight plot rows and, thus, locate the plot center. Multi-threshold and watershed segmentation were then combined to separate plants, ground, and sub-plot within plots. Algorithms of local maxima and pixel resizing with surface tension parameters were used to detect the centers of sub-plots. A total of 3489 reference data points was collected on 30 random plots for accuracy assessment. It was found that all plots and sub-plots were successfully extracted with an overall plot extraction accuracy of 92%. Our methodology addressed some common issues related to plot segmentation, such as plot alignment and overlapping canopies in the field experiments. The ability to segment and extract phenometric information at the sub-plot level provides opportunities to improve the precision of image-based phenotypic measurements at field-scale.

Harnessing Machine Learning Techniques for Mapping Aquaculture Waterbodies in Bangladesh

Aquaculture in Bangladesh has grown dramatically in an unplanned manner in the past few decades, becoming a major contributor to the rural economy in many parts of the country. National systems for the collection of statistics have been unable to keep pace with these rapid changes, and more accurate, up to date information is needed to inform policymakers. Using Sentinel-2 top of atmosphere reflectance data within Google Earth Engine, we proposed six different strategies for improving fishpond detection as the existing techniques seem unreliable. These techniques include: (1) identification of the best time period for image collection, (2) testing the buffer size for threshold optimization, (3) determining the best combination of image reducer and water-identifying indices, (4) introduction of a convolution filter to enhance edge-detection, (5) evaluating the impact of ground truthing data on machine learning algorithm training, and (6) identifying the best machine learning classifier. Each enhancement builds on the previous one to develop a comprehensive improvement strategy called the enhanced method for fishpond detection. We compared the results of each improvement strategy to known ground truthing fishponds as the metric of success. For machine learning classifiers, we compared the precision, recall, and F1 score to determine the quality of results. Among four machine learning methods studied here, the classification and regression trees performed the best with a precision of 0.738, recall of 0.827, and F1 score of 0.780. Overall, the proposed strategies enhanced fishpond area detection in all districts within the study area.

An explainable CNN approach for medical codes prediction from clinical text

Background Clinical notes are unstructured text documents generated by clinicians during patient encounters, generally are annotated with International Classification of Diseases (ICD) codes, which give formatted information about the diagnosis and treatment. ICD code has shown its potentials in many fields, but manual coding is labor-intensive and error-prone, lead to researches of automatic coding. Two specific challenges of this task are (1) given an annotated clinical notes, the reasons behind specific diagnoses and treatments are implicit; (2) explainability is important for practical automatic coding method, the method should not only explain its prediction output but also have explainable internal mechanics. This study aims to develop an explainable CNN approach to address these two challenges. Method Our key idea is that for the automatic ICD coding task, the presence of informative snippets in the clinical text that correlated with each code plays an important role in the prediction of codes, and an informative snippet can be considered as a local and low-level feature. We infer that there exists a correspondence between a convolution filter and a local and low-level feature. Base on the inference, we come up with the Shallow and Wide Attention convolutional Mechanism (SWAM) to improve the CNN-based models’ ability to learn local and low-level features for each label. Results We evaluate our approach on MIMIC-III, an open-access dataset of ICU medical records. Our approach substantially outperforms previous results on top-50 medical code prediction on MIMIC-III dataset, the precision of the worst-performing 10% labels in previous works is increased from 0% to 53% on average. We attribute this improvement to SWAM, by which the wide architecture with attention mechanism gives the model ability to more extensively learn the unique features of different codes, and we prove it by an ablation experiment. Besides, we perform manual analysis of the performance imbalance between different codes, and preliminary conclude the characteristics that determine the difficulty of learning specific codes. Conclusions Our main contributions can be summarized into the following three: (1) We present local and low-level features, a.k.a. informative snippets play an important role in the automatic ICD coding task, and the informative snippets extracted from the clinical text provide explanations for each code. (2) We propose that there exists a correspondence between a convolution filter and a local and low-level feature. A combination of wide and shallow convolutional layer and attention layer can help the CNN-based models better learn local and low-level features. (3) We improved the precision of the worst-performing 10% labels from 0 to 53% on average.

A model of octopus epidermis pattern mimicry mechanisms using inverse operation of the Turing reaction model

Many cephalopods such as octopi and squid can purposefully and rapidly change their skin color. Furthermore, it is widely known that some octopi have the ability to rapidly change the color and unevenness of their skin to mimic their surroundings. However, there has been little research published on the mechanisms by which an octopus recognizes its surrounding landscape and changes its skin pattern. We are unaware of any hypothetical model that explains this mimicry mechanism to date. In this study, the mechanism of octopus skin pattern change was assumed to be based on the Turing pattern model. Here, pattern formation using the Turing model was realized using an equivalent filter calculation model and a cellular automaton instead of directly solving the differential equations. It was shown that this model can create various patterns using two feature parameters. Furthermore, for visual recognition where two features are extracted from the Turing pattern image, a method that requires minimal calculation using the characteristics of the cellular Turing pattern model is proposed. These two calculations can be expressed in the same mathematical frame based on the cellular automaton model using a convolution filter. As a result, a model that is capable of extracting features from patterns and reconstructing those patterns rapidly can be created. This represents a basic model of the mimicry mechanism of octopi. Further, this study demonstrates the potential for creating a model with minimal learning calculation for application to machine learning.

MSB R-CNN: A Multi-Stage Balanced Defect Detection Network

Deep learning networks are applied for defect detection, among which Cascade R-CNN is a multi-stage object detection network and is state of the art in terms of accuracy and efficiency. However, it is still a challenge for Cascade R-CNN to deal with complex and diverse defects, as the widely varied shapes of defects lead to inefficiency for the traditional convolution filter to extract features. Additionally, the imbalance in features, losses and samples cause lower accuracy. To address the above challenges, this paper proposes a multi-stage balanced R-CNN (MSB R-CNN) for defect detection based on Cascade R-CNN. Firstly, deformable convolution is adopted in different stages of the backbone network to improve its adaptability to the varying shapes of the defect. Then, the features obtained by the backbone network are refined and enhanced by the balanced feature pyramid. To overcome the imbalance of classification and regression loss, the balanced L1 loss is applied at different stages to correct it. Finally, for the sample selection, the interaction of union (IoU) balanced sampler and the online hard example mining (OHEM) sampler are combined at different stages to make the sampling more reasonable, which can bring a better accuracy and convergence effect to the model. The results of our experiments on the DAGM2007 dataset has shown that our network (MSB R-CNN) can achieve a mean average precision (mAP) of 67.5%, an increase of 1.5% mAP, compared to Cascade R-CNN.