Small Convolutional Neural Network Research Articles

Deep Learning and Minimally Invasive Endoscopy: Panendoscopic Detection of Pleomorphic Lesions

Introduction: Capsule endoscopy (CE) is a minimally invasive exam suitable of panendoscopic evaluation of the gastrointestinal (GI) tract. Nevertheless, CE is time-consuming with suboptimal diagnostic yield in the upper GI tract. Convolutional neural networks (CNN) are human brain architecture-based models suitable for image analysis. However, there is no study about their role in capsule panendoscopy. Methods: Our group developed an artificial intelligence (AI) model for panendoscopic automatic detection of pleomorphic lesions (namely vascular lesions, protuberant lesions, hematic residues, ulcers, and erosions). 355,110 images (6,977 esophageal, 12,918 gastric, 258,443 small bowel, 76,772 colonic) from eight different CE and colon CE (CCE) devices were divided into a training and validation dataset in a patient split design. The model classification was compared to three CE experts’ classification. The model’s performance was evaluated by its sensitivity, specificity, accuracy, positive predictive value, negative predictive value, and area under the precision-recall curve. Results: The binary esophagus CNN had a diagnostic accuracy for pleomorphic lesions of 83.6%. The binary gastric CNN identified pleomorphic lesions with a 96.6% accuracy. The undenary small bowel CNN distinguished pleomorphic lesions with different hemorrhagic potentials with 97.6% accuracy. The trinary colonic CNN (detection and differentiation of normal mucosa, pleomorphic lesions, and hematic residues) had 94.9% global accuracy. Discussion/Conclusion: We developed the first AI model for panendoscopic automatic detection of pleomorphic lesions in both CE and CCE from multiple brands, solving a critical interoperability technological challenge. Deep learning-based tools may change the landscape of minimally invasive capsule panendoscopy.

A minimalistic approach to classifying Alzheimer’s disease using simple and extremely small convolutional neural networks

Background:There is a broad interest in deploying deep learning-based classification algorithms to identify individuals with Alzheimer’s disease (AD) from healthy controls (HC) based on neuroimaging data, such as T1-weighted Magnetic Resonance Imaging (MRI). The goal of the current study is to investigate whether modern, flexible architectures such as EfficientNet provide any performance boost over more standard architectures. Methods:MRI data was sourced from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and processed with a minimal preprocessing pipeline. Among the various architectures tested, the minimal 3D convolutional neural network SFCN stood out, composed solely of 3x3x3 convolution, batch normalization, ReLU, and max-pooling. We also examined the influence of scale on performance, testing SFCN versions with trainable parameters ranging from 720 up to 2.9 million. Results:SFCN achieves a test ROC AUC of 96.0% while EfficientNet got an ROC AUC of 94.9 %. SFCN retained high performance down to 720 trainable parameters, achieving an ROC AUC of 91.4%. Comparison with existing methods:The SFCN is compared to DenseNet and EfficientNet as well as the results of other publications in the field. Conclusions:The results indicate that using the minimal 3D convolutional neural network SFCN with a minimal preprocessing pipeline can achieve competitive performance in AD classification, challenging the necessity of employing more complex architectures with a larger number of parameters. This finding supports the efficiency of simpler deep learning models for neuroimaging-based AD diagnosis, potentially aiding in better understanding and diagnosing Alzheimer’s disease.

Evaluation of Deformable Convolution: An Investigation in Image and Video Classification

Convolutional Neural Networks (CNNs) present drawbacks for modeling geometric transformations, caused by the convolution operation’s locality. Deformable convolution (DCON) is a mechanism that solves these drawbacks and improves the robustness. In this study, we clarify the optimal way to replace the standard convolution with its deformable counterpart in a CNN model. To this end, we conducted several experiments using DCONs applied in the layers that conform a small four-layer CNN model and on the four-layers of several ResNets with depths 18, 34, 50, and 101. The models were tested in binary balanced classes with 2D and 3D data. If DCON is used on the first layers of the proposal of model, the computational resources will tend to increase and produce bigger misclassification than the standard CNN. However, if the DCON is used at the end layers, the quantity of Flops will decrease, and the classification accuracy will improve by up to 20% about the base model. Moreover, it gains robustness because it can adapt to the object of interest. Also, the best kernel size of the DCON is three. With these results, we propose a guideline and contribute to understanding the impact of DCON on the robustness of CNNs.

From Dataset Creation to Defect Detection: A Proposed Procedure for a Custom CNN Approach for Polishing Applications on Low-Performance PCs

This study focuses on training a custom, small Convolutional Neural Network (CNN) using a limited dataset through data augmentation that is aimed at developing weights for subsequent fine-tuning on specific defects, namely improperly polished aluminum surfaces. The objective is to adapt the network for use in computationally restricted environments. The methodology involves using two computers—a low-performance PC for network creation and initial testing and a more powerful PC for network training using the Darknet framework—after which the network is transferred back to the initial low-performance PC. The results demonstrate that the custom lightweight network suited for a low-performance PC effectively performs object detection under the described conditions. These findings suggest that using tailored lightweight networks for recognizing specific types of defects is feasible and warrants further investigation to enhance the industrial defect detection processes in limited computational settings. This approach highlights the potential for deploying AI-driven quality control in environments with constrained hardware capabilities.

Emergency sign language recognition from variant of convolutional neural network (CNN) and long short term memory (LSTM) models

Sign language is the primary communication tool used by the deaf community and people with speaking difficulties, especially during emergencies. Numerous deep learning models have been proposed to solve the sign language recognition problem. Recently. Bidirectional LSTM (BLSTM) has been proposed and used in replacement of Long Short-Term Memory (LSTM) as it may improve learning long-team dependencies as well as increase the accuracy of the model. However, there needs to be more comparison for the performance of LSTM and BLSTM in LRCN model architecture in sign language interpretation applications. Therefore, this study focused on the dense analysis of the LRCN model, including 1) training the CNN from scratch and 2) modeling with pre-trained CNN, VGG-19, and ResNet50. Other than that, the ConvLSTM model, a special variant of LSTM designed for video input, has also been modeled and compared with the LRCN in representing emergency sign language recognition. Within LRCN variants, the performance of a small CNN network was compared with pre-trained VGG-19 and ResNet50V2. A dataset of emergency Indian Sign Language with eight classes is used to train the models. The model with the best performance is the VGG-19 + LSTM model, with a testing accuracy of 96.39%. Small LRCN networks, which are 5 CNN subunits + LSTM and 4 CNN subunits + BLSTM, have 95.18% testing accuracy. This performance is on par with our best-proposed model, VGG + LSTM. By incorporating bidirectional LSTM (BLSTM) into deep learning models, the ability to understand long-term dependencies can be improved. This can enhance accuracy in reading sign language, leading to more effective communication during emergencies.

Siamese Networks for Clinically Relevant Bacteria Classification Based on Raman Spectroscopy.

Identifying bacterial strains is essential in microbiology for various practical applications, such as disease diagnosis and quality monitoring of food and water. Classical machine learning algorithms have been utilized to identify bacteria based on their Raman spectra. However, convolutional neural networks (CNNs) offer higher classification accuracy, but they require extensive training sets and retraining of previous untrained class targets can be costly and time-consuming. Siamese networks have emerged as a promising solution. They are composed of two CNNs with the same structure and a final network that acts as a distance metric, converting the classification problem into a similarity problem. Classical machine learning approaches, shallow and deep CNNs, and two Siamese network variants were tailored and tested on Raman spectral datasets of bacteria. The methods were evaluated based on mean sensitivity, training time, prediction time, and the number of parameters. In this comparison, Siamese-model2 achieved the highest mean sensitivity of 83.61 ± 4.73 and demonstrated remarkable performance in handling unbalanced and limited data scenarios, achieving a prediction accuracy of 73%. Therefore, the choice of model depends on the specific trade-off between accuracy, (prediction/training) time, and resources for the particular application. Classical machine learning models and shallow CNN models may be more suitable if time and computational resources are a concern. Siamese networks are a good choice for small datasets and CNN for extensive data.

Small Convolutional Neural Network Trainer Designed through Transfer Learning in Dessert Classification

<p>This paper established a Convolutional Neural Network (CNN) with the concept of transfer learning, and combined the main feature analysis calculation and clustering algorithm to further demonstrate the superiority of the proposed small CNN trainer in the identification of traditional Kinmen desserts. Food dessert identification methods directly use skin texture, color, shape, and other features as the basis. This paper effectively extracted image features of an object by the small CNN trainer and classified the featured dataset into the right food categories. It was able to complete classification quickly and also achieved high-precision classification results. Different types of Kinmen desserts were identified through a multi-layer training cycle. A total of 100 training images for each of the 10 food categories and each image size is converted into a smaller training data set by capturing the important features through the CNN trainer. Then, the main features were generated and the dimensions of each food image data were reduced again by using the t-Distributed Stochastic Neighbor Embedding (t-SNE) or Principal Component Analysis (PCA) methods. An individually K-means or k-nearest neighbors (KNN) algorithms efficiently completed the grouping results and in the classified image restoration. The experimental results compared the classifications after the learning cycle of different trainers and showed that the highest accuracy that the appropriated CNN trainer of the proposed methology obtained was up to 99% with a minimum executing time of about 99.37 seconds.</p> <p> </p>

DTV-CNN: Neural network based on depth and thickness views for efficient 3D shape classification

Fast and effective algorithms for deep learning on 3D shapes are keys to innovate mechanical and electronic engineering design workflow. In this paper, an efficient 3D shape to 2D images projection algorithm and a shallow 2.5D convolutional neural network architecture is proposed. A smaller convolutional neural network (CNN) model is achieved by information enrichment at the preprocessing stage, i.e. 3D geometry is compressed into 2D “thickness view” and “depth view”. Fusing the depth view and thickness view (DTV) from the same projection view into a dual-channel grayscale image, can improve information locality for geometry and topology feature extraction. This approach bridges the gap between mature image deep learning technologies to the applications of 3D shape. Enhanced by several essential scalar geometry properties and only 3 projection views, a mixed CNN and multiple linear parameter (MLP) neural network model achives a validation accuracy of 92 % for ModelNet10 mesh-based dataset, while the training time is one order of magnitude less than the original multi-view CNN approach. This study also creates new 3D shape datasets from 2 open source CAD projects. Higher validation accuracy is obtained for realistic CAD datasets, i.e. 97 % for FreeCAD's mechanical part library and 95 % for KiCAD electronic part library. The training cost reduces to tens of minutes on a laptop CPU, given the smaller input data size and shallow neural network design. It is expected that this approach can be adapted for other machine learning scenarios involved in CAD geometry.

Parameter optimization in convolutional neural network for improving performance of facial expression recognition

The recognition of similar facial expressions presents a notable challenge, necessitating a focus on the parameters within the fundamental Convolutional Neural Network (CNN) architecture, which serves as a cornerstone in the field of image classification. This research endeavor aims to enhance the model's capacity for facial expression recognition by employing a controlled variable method to examine two specific parameters in a self-designed small CNN: the number of filters and convolutional layers. More specifically, while the filters were fixed at 3, the layers varied from 3 to 6 to 9. Similarly, as the number of the filters rose to 6, the number of the layers also incremented from 3 to 6 to 9. Furthermore, while the number of the filters reached 12, the number of the layers went from 3 to 6 to 9 too. Finally, with the filters increasing to 24, the layers rose from 3 to 6 to 9 as well. Experimental results indicate that both increasing the number of filters and convolutional layers can increment the performance of model in facial expression recognition. Furthermore, increasing the number of filters can exert a more prominent influence on improving the accuracy of facial expression recognition.

Improving Nighttime Driving-Scene Segmentation via Dual Image-Adaptive Learnable Filters

Semantic segmentation on driving-scene images is vital for autonomous driving. Although encouraging performance has been achieved on daytime images, the performance on nighttime images are less satisfactory due to the insufficient exposure and the lack of labeled data. To address these issues, we present an add-on module called dual image-adaptive learnable filters (DIAL-Filters) to improve the semantic segmentation in nighttime driving conditions, aiming at exploiting the intrinsic features of driving-scene images under different illuminations. DIAL-Filters consist of two parts, including an image-adaptive processing module (IAPM) and a learnable guided filter (LGF). With DIAL-Filters, we design both unsupervised and supervised frameworks for nighttime driving-scene segmentation, which can be trained in an end-to-end manner. Specifically, the IAPM module consists of a small convolutional neural network with a set of differentiable image filters, where each image can be adaptively enhanced for better segmentation with respect to the different illuminations. The LGF is employed to enhance the output of segmentation network to get the final segmentation result. The DIAL-Filters are light-weight and efficient and they can be readily applied for both daytime and nighttime images. Our experiments show that DAIL-Filters can significantly improve the supervised segmentation performance on ACDC_Night and NightCity datasets, while it demonstrates the state-of-the-art performance on unsupervised nighttime semantic segmentation on Dark Zurich and Nighttime Driving testbeds. Codes and models are available at https://github.com/wenyyu/IA-Seg.

A Modified LeNet CNN for Breast Cancer Diagnosis in Ultrasound Images.

Convolutional neural networks (CNNs) have been extensively utilized in medical image processing to automatically extract meaningful features and classify various medical conditions, enabling faster and more accurate diagnoses. In this paper, LeNet, a classic CNN architecture, has been successfully applied to breast cancer data analysis. It demonstrates its ability to extract discriminative features and classify malignant and benign tumors with high accuracy, thereby supporting early detection and diagnosis of breast cancer. LeNet with corrected Rectified Linear Unit (ReLU), a modification of the traditional ReLU activation function, has been found to improve the performance of LeNet in breast cancer data analysis tasks via addressing the "dying ReLU" problem and enhancing the discriminative power of the extracted features. This has led to more accurate, reliable breast cancer detection and diagnosis and improved patient outcomes. Batch normalization improves the performance and training stability of small and shallow CNN architecture like LeNet. It helps to mitigate the effects of internal covariate shift, which refers to the change in the distribution of network activations during training. This classifier will lessen the overfitting problem and reduce the running time. The designed classifier is evaluated against the benchmarking deep learning models, proving that this has produced a higher recognition rate. The accuracy of the breast image recognition rate is 89.91%. This model will achieve better performance in segmentation, feature extraction, classification, and breast cancer tumor detection.

Accelerated respiratory-resolved 4D-MRI with separable spatio-temporal neural networks.

Respiratory-resolved four-dimensional magnetic resonance imaging (4D-MRI) provides essential motion information for accurate radiation treatments of mobile tumors. However, obtaining high-quality 4D-MRI suffers from long acquisition and reconstruction times. To develop a deep learning architecture to quickly acquire and reconstruct high-quality 4D-MRI, enabling accurate motion quantification for MRI-guided radiotherapy (MRIgRT). A small convolutional neural network called MODEST is proposed to reconstruct 4D-MRI by performing a spatial and temporal decomposition, omitting the need for 4D convolutions to use all the spatio-temporal information present in 4D-MRI. This network is trained on undersampled 4D-MRI after respiratory binning to reconstruct high-quality 4D-MRI obtained by compressed sensing reconstruction. The network is trained, validated, and tested on 4D-MRI of 28 lung cancer patients acquired with a T1-weighted golden-angle radial stack-of-stars (GA-SOS) sequence. The 4D-MRI of 18, 5, and 5 patients were used for training, validation, and testing. Network performances are evaluated on image quality measured by the structural similarity index (SSIM) and motion consistency by comparing the position of the lung-liver interface on undersampled 4D-MRI before and after respiratory binning. The network is compared to conventional architectures such as a U-Net, which has 30 times more trainableparameters. MODEST can reconstruct high-quality 4D-MRI with higher image quality than a U-Net, despite a thirty-fold reduction in trainable parameters. High-quality 4D-MRI can be obtained using MODEST in approximately 2.5 min, including acquisition, processing, andreconstruction. High-quality accelerated 4D-MRI can be obtained using MODEST, which is particularly interesting forMRIgRT.

Artificial visual perception neural system using a solution-processable MoS2-based in-memory light sensor

Optoelectronic devices are advantageous in in-memory light sensing for visual information processing, recognition, and storage in an energy-efficient manner. Recently, in-memory light sensors have been proposed to improve the energy, area, and time efficiencies of neuromorphic computing systems. This study is primarily focused on the development of a single sensing-storage-processing node based on a two-terminal solution-processable MoS2 metal–oxide–semiconductor (MOS) charge-trapping memory structure—the basic structure for charge-coupled devices (CCD)—and showing its suitability for in-memory light sensing and artificial visual perception. The memory window of the device increased from 2.8 V to more than 6 V when the device was irradiated with optical lights of different wavelengths during the program operation. Furthermore, the charge retention capability of the device at a high temperature (100 °C) was enhanced from 36 to 64% when exposed to a light wavelength of 400 nm. The larger shift in the threshold voltage with an increasing operating voltage confirmed that more charges were trapped at the Al2O3/MoS2 interface and in the MoS2 layer. A small convolutional neural network was proposed to measure the optical sensing and electrical programming abilities of the device. The array simulation received optical images transmitted using a blue light wavelength and performed inference computation to process and recognize the images with 91% accuracy. This study is a significant step toward the development of optoelectronic MOS memory devices for neuromorphic visual perception, adaptive parallel processing networks for in-memory light sensing, and smart CCD cameras with artificial visual perception capabilities.

Small Space Target Detection Based on a Convolutional Neural Network and Guidance Information

Although space targets have different shapes, sizes and intensities, their distributions share certain commonalities. However, it is difficult to summarize a generalized distribution function for space targets. Moreover, most of the existing methods based on deep learning are not suitable to use directly because of the size of targets and the cost of manual labeling for a full image. In this paper, we proposed a pattern for space target detection based on a convolutional neural network (CNN) to learn essential features of the targets from data. In the processing stage, the background is estimated and removed. Then, image techniques are used to search and process region proposals. Different sizes of region proposals are recognized by a discriminator, which is built upon a small CNN trained with the data of several specific targets. Finally, a non-maximum suppression (NMS) operation is used to remove redundant targets. In the network structure, to further enhance the influence of the effective area, the parameters calculated from the center region of the input are utilized as guidance information and added to the features before the full connection. Moreover, the bias loss is applied to increase the weights of unique features. The experimental results demonstrate the outstanding performance of the proposed method in terms of the number of detected targets, accuracy rate and false alarm rate compared with baseline methods. In particular, the proposed method has a simple network structure and a lower computational cost which can be further promoted and implemented in actual engineering.

EVlncRNA-Dpred: improved prediction of experimentally validated lncRNAs by deep learning.

Long non-coding RNAs (lncRNAs) played essential roles in nearly every biological process and disease. Many algorithms were developed to distinguish lncRNAs from mRNAs in transcriptomic data and facilitated discoveries of more than 600000 of lncRNAs. However, only a tiny fraction (<1%) of lncRNA transcripts (~4000) were further validated by low-throughput experiments (EVlncRNAs). Given the cost and labor-intensive nature of experimental validations, it is necessary to develop computational tools to prioritize those potentially functional lncRNAs because many lncRNAs from high-throughput sequencing (HTlncRNAs) could be resulted from transcriptional noises. Here, we employed deep learning algorithms to separate EVlncRNAs from HTlncRNAs and mRNAs. For overcoming the challenge of small datasets, we employed a three-layer deep-learning neural network (DNN) with a K-mer feature as the input and a small convolutional neural network (CNN) with one-hot encoding as the input. Three separate models were trained for human (h), mouse (m) and plant (p), respectively. The final concatenated models (EVlncRNA-Dpred (h), EVlncRNA-Dpred (m) and EVlncRNA-Dpred (p)) provided substantial improvement over a previous model based on support-vector-machines (EVlncRNA-pred). For example, EVlncRNA-Dpred (h) achieved 0.896 for the area under receiver-operating characteristic curve, compared with 0.582 given by sequence-based EVlncRNA-pred model. The models developed here should be useful for screening lncRNA transcripts for experimental validations. EVlncRNA-Dpred is available as a web server at https://www.sdklab-biophysics-dzu.net/EVlncRNA-Dpred/index.html, and the data and source code can be freely available along with the web server.

ReSTiNet: An Efficient Deep Learning Approach to Improve Human Detection Accuracy

Human detection is an important task in computer vision. It is one of the most important tasks in global security and safety monitoring. In recent days, Deep Learning has improved human detection technology. Despite modern techniques, there are very few optimal techniques to construct networks with a small size, deep architecture, and fast training time while maintaining accuracy. ReSTiNet is a novel small convolutional neural network that overcomes the problems of network size, detection speed, and accuracy. The developed ReSTiNet contains fire modules by evaluating their number and position in the network to minimize the model parameters and network size. To improve the detection speed and accuracy of ReSTiNet, the residual block within the fire modules is carefully designed to increase the feature propagation and maximize the information flow in the network. The developed approach compresses the well-known Tiny-YOLO architecture while improving the following features: (i) small model size, (ii) faster detection speed, (iii) resolution of overfitting, and (iv) better performance than other compact networks such as SqueezeNet and MobileNet in terms of mAP on the Pascal VOC and MS COCO datasets.ReSTiNet is 10.7 MB, five times smaller than Tiny-YOLO. On Tesla k80, mAP is 27.3% for MS COCO and 63.74% for PASCAL VOC. The validation of the proposed ReSTiNet model has been done on INRIA person dataset using the Tesla K80.•All the necessary steps, algorithms, and mathematical formulas for building the net- work are provided.•The network is small in size but has a faster detection speed with high accuracy.

A Novel Deep Learning-Based Pose Estimation Method for Robotic Grasping of Axisymmetric Bodies in Industrial Stacked Scenarios

A vision-based intelligent robotic grasping system is essential for realizing unmanned operations in industrial manufacturing, and pose estimation plays an import role in this system. In this study, deep learning was used to obtain the 6D pose of an axisymmetric body which was optimal for robotic grasping in industrial stacked scenarios. We propose a method to obtain the 6D pose of an axisymmetric body by detecting the pre-defined keypoints on the side surface. To realize this method and solve other challenges in industrial stacked scenarios, we propose a multitask real-time convolutional neural network (CNN), named Key-Yolact, which involves object detection, instance segmentation, and multiobject 2D keypoint detection. A small CNN as a decision-making subsystem was designed to score multiple predictions of Key-Yolact, and the body with the highest score is considered the best for grasping. Experiments on a self-built stacked dataset showed that Key-Yolact has a practical tradeoff between inference speed and precision. The inference speed of Key-Yolact is higher by 10 FPS, whereas its precision is decreased by only 7% when compared with the classical multitask Keypoint R-CNN. Robotic grasping experiments showed that the proposed design is effective and can be directly applied to industrial scenarios.

Patch-wise weed coarse segmentation mask from aerial imagery of sesame crop

Sesame (Sesamum indicum L.) is an important commercial food crop and there are numerous weed species that reduce their yield. Drone-assisted autonomous aerial herbicide spraying on sesame fields intends to prevent weeds from limiting agricultural productivity. Additionally, agrochemical application quantity and human health are safeguarded using aerial spraying. Autonomous systems would require characteristics to differentiate between crop, weed, and background for accurate and selective spray application. Datasets from aerial sesame fields have been gathered for this study to extract weed and crop profiles in the area. Agrocam, a camera sensor made to track crop health, was used to record the dataset. We have created an intelligent SesameWeedNet-2 convolutional neural network (CNN) and a patch image-based classification method. The eight convolutional layers in the tiny model allow it to operate more quickly and precisely on small patch images. Our approach divides photos into small patch images of 31 × 31 pixels. After that, the proposed small CNN is trained and tested on these small patch images. Two critical components of our methodology are dataset grouping and patch-based model ensemble. Our algorithm suggests classifying the dataset based on the vegetation visible in the small patch images to improve categorization outcomes. Our suggested strategy has attained an accuracy of up to 96.7 %. We have made comparisons among our suggested patch-based technique with pixel-level semantic segmentation. The patch-based method wins for time efficiency even though the pixel-level method provides a fine outline of the crop and weed classes but is computationally expensive. Our approach has been put to the test in both dry and wet soil conditions, as well as at various growth phases. We are aware of no prior attempts to categorize and treat crops and weeds in sesame fields at the post-emergence stage. A complete deep learning-based patch-based technique and an aerial sesame-weed dataset contribute to this study to categorize weeds in sesame fields under varying lighting circumstances.

A Garbage Classification Method Based on a Small Convolution Neural Network

To improve the efficiency of social garbage classification, a garbage classification method based on a small convolutional neural network (CNN) is proposed in this paper. For low accuracy caused by light and shadow interference, an adaptive image-brightening algorithm is developed to average the brightness of the background in the image preprocessing stage, and a threshold replacement method is used to reduce shadow noise. Then, the Canny operator is used to assist in cropping the blank background in the image. For debugging low efficiency caused by the complex network, the neural network is optimized based on the MLH-CNN model to make its results simpler and equally efficient. Experimental results show the preprocessing in this study can improve the accuracy of model garbage classification. The CNN model in this study can achieve an accuracy of 96.77% on the self-built dataset and 93.72% on the TrashNet dataset, which is higher than the 92.6% accuracy of the MLC-CNN model. The network optimizer can also enhance the classification ability of the network model using the Adamax optimization algorithm based on Adam variants. In this paper, the network model derived from training is combined with the host computer software to design a garbage detection page so the model has a wider range of uses, which has a good effect on promoting the development of social environmental protection and improving residents’ awareness of environmental protection.

Kernel Density Estimation and Convolutional Neural Networks for the Recognition of Multi-Font Numbered Musical Notation

Optical music recognition (OMR) refers to converting musical scores into digitized information using electronics. In recent years, few types of OMR research have involved numbered musical notation (NMN). The existing NMN recognition algorithm is difficult to deal with because the numbered notation font is changing. In this paper, we made a multi-font NMN dataset. Using the presented dataset, we use kernel density estimation with proposed bar line criteria to measure the relative height of symbols, and an accurate separation of melody lines and lyrics lines in musical notation is achieved. Furthermore, we develop a structurally improved convolutional neural network (CNN) to classify the symbols in melody lines. The proposed neural network performs hierarchical processing of melody lines according to the symbol arrangement rules of NMN and contains three parallel small CNNs called Arcnet, Notenet and Linenet. Each of them adds a spatial pyramid pooling layer to adapt to the diversity of symbol sizes and styles. The experimental results show that our algorithm can accurately detect melody lines. Taking the average accuracy rate of identifying various symbols as the recognition rate, the improved neural networks reach a recognition rate of 95.5%, which is 8.5% higher than the traditional convolutional neural networks. Through audio comparison and evaluation experiments, we find that the generated audio maintains a high similarity to the original audio of the NMN.