Pure Convolutional Neural Networks Research Articles

InViTMixup: plant disease classification using convolutional vision transformer with Mixup augmentation

ABSTRACT Computer vision-based research is carried out to enhance the crop yield with automation solutions. In Tamil Nādu, tomato cultivation is carried out in many districts, including Coimbatore, and its contribution to the Indian economy, i.e. GDP growth, is significant. This paper proposes InViT Mixup, a novel approach toward image classification of diseased tomato leaves. The effectiveness of convolutional neural network (CNN) in obtaining spatial information and the attention mechanism of vision transformers, along with Mixup data augmentation, are combined, and the model’s generalization in predicting the diseased tomato based on leaf symptoms is evaluated. The proposed convolutional transformer model improvised the classification results of tomato disease in terms of accuracy and training speed compared to pure CNN or transformer-based models. In the proposed work, 10 tomato leaf disease classes from the PlantVillage dataset are used. Top-1 accuracy of 93.5% and Top-5 accuracy of 99.7% are achieved.

Pure Vision Transformer (CT-ViT) with Noise2Neighbors Interpolation for Low-Dose CT Image Denoising.

Convolutional neural networks (CNN) have been used for a wide variety of deep learning applications, especially in computer vision. For medical image processing, researchers have identified certain challenges associated with CNNs. These challenges encompass the generation of less informative features, limitations in capturing both high and low-frequency information within feature maps, and the computational cost incurred when enhancing receptive fields by deepening the network. Transformers have emerged as an approach aiming to address and overcome these specific limitations of CNNs in the context of medical image analysis. Preservation of all spatial details of medical images is necessary to ensure accurate patient diagnosis. Hence, this research introduced the use of a pure Vision Transformer (ViT) for a denoising artificial neural network for medical image processing specifically for low-dose computed tomography (LDCT) image denoising. The proposed model follows a U-Net framework that contains ViT modules with the integration of Noise2Neighbor (N2N) interpolation operation. Five different datasets containing LDCT and normal-dose CT (NDCT) image pairs were used to carry out this experiment. To test the efficacy of the proposed model, this experiment includes comparisons between the quantitative and visual results among CNN-based (BM3D, RED-CNN, DRL-E-MP), hybrid CNN-ViT-based (TED-Net), and the proposed pure ViT-based denoising model. The findings of this study showed that there is about 15-20% increase in SSIM and PSNR when using self-attention transformers than using the typical pure CNN. Visual results also showed improvements especially when it comes to showing fine structural details of CT images.

MSCT-UNET: multi-scale contrastive transformer within U-shaped network for medical image segmentation

Objective. Automatic mutli-organ segmentation from anotomical images is essential in disease diagnosis and treatment planning. The U-shaped neural network with encoder–decoder has achieved great success in various segmentation tasks. However, a pure convolutional neural network (CNN) is not suitable for modeling long-range relations due to limited receptive fields, and a pure transformer is not good at capturing pixel-level features. Approach. We propose a new hybrid network named MSCT-UNET which fuses CNN features with transformer features at multi-scale and introduces multi-task contrastive learning to improve the segmentation performance. Specifically, the multi-scale low-level features extracted from CNN are further encoded through several transformers to build hierarchical global contexts. Then the cross fusion block fuses the low-level and high-level features in different directions. The deep-fused features are flowed back to the CNN and transformer branch for the next scale fusion. We introduce multi-task contrastive learning including a self-supervised global contrast learning and a supervised local contrast learning into MSCT-UNET. We also make the decoder stronger by using a transformer to better restore the segmentation map. Results. Evaluation results on ACDC, Synapase and BraTS datasets demonstrate the improved performance over other methods compared. Ablation study results prove the effectiveness of our major innovations. Significance. The hybrid encoder of MSCT-UNET can capture multi-scale long-range dependencies and fine-grained detail features at the same time. The cross fusion block can fuse these features deeply. The multi-task contrastive learning of MSCT-UNET can strengthen the representation ability of the encoder and jointly optimize the networks. The source code is publicly available at: https://github.com/msctunet/MSCT_UNET.git.

Cloth Interactive Transformer for Virtual Try-On

The 2D image-based virtual try-on has aroused increased interest from the multimedia and computer vision fields due to its enormous commercial value. Nevertheless, most existing image-based virtual try-on approaches directly combine the person-identity representation and the in-shop clothing items without taking their mutual correlations into consideration. Moreover, these methods are commonly established on pure convolutional neural networks (CNNs) architectures which are not simple to capture the long-range correlations among the input pixels. As a result, it generally results in inconsistent results. To alleviate these issues, in this article, we propose a novel two-stage cloth interactive transformer (CIT) method for the virtual try-on task. During the first stage, we design a CIT matching block, aiming at precisely capturing the long-range correlations between the cloth-agnostic person information and the in-shop cloth information. Consequently, it makes the warped in-shop clothing items look more natural in appearance. In the second stage, we put forth a CIT reasoning block for establishing global mutual interactive dependencies among person representation, the warped clothing item, and the corresponding warped cloth mask. The empirical results, based on mutual dependencies, demonstrate that the final try-on results are more realistic. Substantial empirical results on a public fashion dataset illustrate that the suggested CIT attains competitive virtual try-on performance.

Endoscopic image classification algorithm based on Poolformer.

Image desmoking is a significant aspect of endoscopic image processing, effectively mitigating visual field obstructions without the need for additional surgical interventions. However, current smoke removal techniques tend to apply comprehensive video enhancement to all frames, encompassing both smoke-free and smoke-affected images, which not only escalates computational costs but also introduces potential noise during the enhancement of smoke-free images. In response to this challenge, this paper introduces an approach for classifying images that contain surgical smoke within endoscopic scenes. This classification method provides crucial target frame information for enhancing surgical smoke removal, improving the scientific robustness, and enhancing the real-time processing capabilities of image-based smoke removal method. The proposed endoscopic smoke image classification algorithm based on the improved Poolformer model, augments the model's capacity for endoscopic image feature extraction. This enhancement is achieved by transforming the Token Mixer within the encoder into a multi-branch structure akin to ConvNeXt, a pure convolutional neural network. Moreover, the conversion to a single-path topology during the prediction phase elevates processing speed. Experiments use the endoscopic dataset sourced from the Hamlyn Centre Laparoscopic/Endoscopic Video Dataset, augmented by Blender software rendering. The dataset comprises 3,800 training images and 1,200 test images, distributed in a 4:1 ratio of smoke-free to smoke-containing images. The outcomes affirm the superior performance of this paper's approach across multiple parameters. Comparative assessments against existing models, such as mobilenet_v3, efficientnet_b7, and ViT-B/16, substantiate that the proposed method excels in accuracy, sensitivity, and inference speed. Notably, when contrasted with the Poolformer_s12 network, the proposed method achieves a 2.3% enhancement in accuracy, an 8.2% boost in sensitivity, while incurring a mere 6.4 frames per second reduction in processing speed, maintaining 87 frames per second. The results authenticate the improved performance of the refined Poolformer model in endoscopic smoke image classification tasks. This advancement presents a lightweight yet effective solution for the automatic detection of smoke-containing images in endoscopy. This approach strikes a balance between the accuracy and real-time processing requirements of endoscopic image analysis, offering valuable insights for targeted desmoking process.

Deep internally connected transformer hashing for image retrieval

Transformer based on self-attention mechanism has made remarkable achievements in natural language processing, which inspired the application research of Transformer in computer vision. The current deep hashing algorithms extract image features through the convolutional neural network (CNN). CNN concentrates on local information, and features lack global dependency information, which has an impact on image retrieval accuracy. To remedy the above defects, this paper proposes deep internally connected Transformer hashing for image retrieval (DICTH). DICTH has designed an improved Transformer block: internally connected Transformer block (ICT). ICT performs an embedded transformation on the feature maps, splices the generated Keys and Queries, to explore the rich context information between Query–Key pairs, and then dynamically encodes through multi-layer convolution to learn the context multi-head self-attention matrix. By combining ICT and ResNet18 to achieve self-attention injection, a long-distance dependency is established in the feature space to make up for the shortcomings of pure CNN in the feature extraction process and guide the algorithm to learn more accurate hash codes. At the same time, in the face of complex label information in big data sets, this paper uses an improved cross-entropy loss function: T-cross-entropy loss, to promote network learning of hash codes with more ability to distinguish between classes. In this paper, a lot of experiments have been conducted on CIFAR10, NUS-WIDE and MS-COCO datasets to verify the performance of DICTH.

Transformer-based ensemble deep learning model for EEG-based emotion recognition

Emotion recognition is one of the most important research directions in the field of brain–computer interface (BCI). However, to conduct electroencephalogram (EEG)-based emotion recognition, there exist difficulties regarding EEG signal processing; moreover, the performance of classification models in this regard is restricted. To counter these issues, the 2022 World Robot Contest successfully held an affective BCI competition, thus promoting the innovation of EEG-based emotion recognition. In this paper, we propose the Transformer-based ensemble (TBEM) deep learning model. TBEM comprises two models: a pure convolutional neural network (CNN) model and a cascaded CNN-Transformer hybrid model. The proposed model won the abovementioned affective BCI competition’s final championship in the 2022 World Robot Contest, demonstrating the effectiveness of the proposed TBEM deep learning model for EEG-based emotion recognition.

Comparative Study of Herbal Leaves Classification using Hybrid of GLCM-SVM and GLCM-CNN

Indonesia is a tropical country with a diverse range of plants that ancient people used for traditional medicines. However, the similarity in shape of the leaves became an obstacle to distinguishing them. Therefore, technological advancements are expected to help identify the herbal leaves to use them right on target according to their efficacy. In this research, image classification of katuk (Sauropus Androgynus) and kelor (Moringa Oleifera) leaves is applied using 3 different algorithms i.e hybrid of Gray Level Co-Occurrence Matrix (GLCM) feature extraction and Support Vector Machine (SVM) implementing 4 kernels namely linear, RBF, polynomial, and sigmoid; hybrid of GLCM and Convolutional Neural Network (CNN); and pure CNN. A dataset of 480 images has been collected with 2 different scenarios, including bright and dark intensities. Based on the result, a hybrid of GLCM and SVM showed the highest accuracy of 96% in the dark intensity test using a linear kernel, while sigmoid obtained the lowest accuracy of 35%. On the other hand, it has been discovered that CNN obtained the highest performance in the bright intensity test with an accuracy of 98%. While in the dark intensity test, a hybrid of GLCM and CNN is superior, obtaining 96% accuracy. In conclusion, CNN is more powerful for image classification with bright intensity. For dark intensity images, both the hybrid of GLCM+SVM (linear) and the hybrid of GLCM+CNN are fairly recommended.

Eff-PCNet: An Efficient Pure CNN Network for Medical Image Classification

With the development of deep learning, convolutional neural networks (CNNs) and Transformer-based methods have become key techniques for medical image classification tasks. However, many current neural network models have problems such as high complexity, a large number of parameters, and large model sizes; such models obtain higher classification accuracy at the expense of lightweight networks. Moreover, such larger-scale models pose a great challenge for practical clinical applications. Meanwhile, Transformer and multi-layer perceptron (MLP) methods have some shortcomings in terms of local modeling capability and high model complexity, and need to be used on larger datasets to show good performance. This makes it difficult to utilize these networks in clinical medicine. Based on this, we propose a lightweight and efficient pure CNN network for medical image classification (Eff-PCNet). On the one hand, we propose a multi-branch multi-scale CNN (M2C) module, which divides the feature map into four parallel branches along the channel dimensions by a certain scale factor and carries out a deep convolution operation using different scale convolution kernels, and this multi-branch multi-scale operation effectively replaces the large kernel convolution. This multi-branch multi-scale operation effectively replaces the large kernel convolution. It reduces the computational cost of the module while fusing the feature information between different channels and thus obtains richer feature information. Finally, the four feature maps are then spliced along the channel dimensions to fuse the multi-scale and multi-dimensional feature information. On the other hand, we introduce the structural reparameterization technique and propose the structural reparameterized CNN (Rep-C) module. Specifically, it utilizes multiple linear operators to generate different feature maps during the training process and fuses all the participants into one through parameter fusion to achieve fast inference while providing a more effective solution for feature reuse. A number of experimental results show that our Eff-PCNet performs better than current methods based on CNN, Transformer, and MLP in the classification of three publicly available medical image datasets. Among them, we achieve 87.4% Acc on the HAM10000 dataset, 91.06% Acc on the SkinCancer dataset, and 97.03% Acc on the Chest-Xray dataset. Meanwhile, our approach achieves a better trade-off between the number of parameters; computation; and other performance metrics as well.

A CatBoost-Based Model for the Intensity Detection of Tropical Cyclones over the Western North Pacific Based on Satellite Cloud Images

A CatBoost-based intelligent tropical cyclone (TC) intensity-detecting model was built to quantify the intensity of TCs over the Western North Pacific (WNP) with the cloud-top brightness temperature (CTBT) data of Fengyun-2F (FY-2F) and Fengyun-2G (FY-2G) and the best-track data of the China Meteorological Administration (CMA-BST) in recent years (2015–2018). The CatBoost-based model was featured with the greedy strategy of combination, the ordering principle in optimizing the possible gradient bias and prediction shift problems, and the oblivious tree in fast scoring. Compared with the previous studies based on the pure convolutional neural network (CNN) models, the CatBoost-based model exhibited better skills in detecting the TC intensity with the root mean square error (RMSE) of 3.74 m s−1. In addition to the three mentioned model features, there are also two reasons for the model’s design. On one hand, the CatBoost-based model used the method of introducing prior physical factors (e.g., the structure and shape of the cloud, deep convections, and background fields) into its training process. On the other hand, the CatBoost-based model expanded the dataset size from 2342 to 13,471 samples through hourly interpolations of the original dataset. Furthermore, this paper investigated the errors of this model in detecting the different categories of TC intensity. The results showed that the deep learning-based TC intensity-detecting model proposed in this paper has systematic biases, namely, the overestimation (underestimation) of intensities in TCs which were weaker (stronger) than at the typhoon level, and the errors of the model in detecting weaker (stronger) TCs were smaller (larger). This implies that more factors than the CTBT should be included to further reduce the errors in detecting strong TCs.

Surrounding-aware representation prediction in Birds-Eye-View using transformers.

Birds-Eye-View (BEV) maps provide an accurate representation of sensory cues present in the surroundings, including dynamic and static elements. Generating a semantic representation of BEV maps can be a challenging task since it relies on object detection and image segmentation. Recent studies have developed Convolutional Neural networks (CNNs) to tackle the underlying challenge. However, current CNN-based models encounter a bottleneck in perceiving subtle nuances of information due to their limited capacity, which constrains the efficiency and accuracy of representation prediction, especially for multi-scale and multi-class elements. To address this issue, we propose novel neural networks for BEV semantic representation prediction that are built upon Transformers without convolution layers in a significantly different way from existing pure CNNs and hybrid architectures that merge CNNs and Transformers. Given a sequence of image frames as input, the proposed neural networks can directly output the BEV maps with per-class probabilities in end-to-end forecasting. The core innovations of the current study contain (1) a new pixel generation method powered by Transformers, (2) a novel algorithm for image-to-BEV transformation, and (3) a novel network for image feature extraction using attention mechanisms. We evaluate the proposed Models performance on two challenging benchmarks, the NuScenes dataset and the Argoverse 3D dataset, and compare it with state-of-the-art methods. Results show that the proposed model outperforms CNNs, achieving a relative improvement of 2.4 and 5.2% on the NuScenes and Argoverse 3D datasets, respectively.

HUMAN GAIT ANALYSIS USING HYBRID CONVOLUTIONAL NEURAL NETWORKS

Human gait analysis is a promising method of researching on human activities like walking or sitting. It reflects the habits of one person and can be observed in any activity that person performs. The patterns in human movements are influenced by many factors, including physiology, social, psychological, and health factors. Differences in limb movements help identify gait patterns, which are often measured using inertial measurement unit sensors (IMU) like gyroscopes and accelerometers placed in various locations throughout the body. This paper analyses the combination of IMU sensors and electromyography sensors (EMG) to improve the identification accuracy of human movements. We propose the hybrid convolutional neural network (CNN) and long short-term memory neuron network (LSTM) for the human gait analysis problem to achieve an accuracy of 0.9418, better than other models including pure CNN models. By using CNN's image classification advancements, we analyse multivariate time series sensor signals by using a sliding window to transform sensor data into image representation and principal component analysis (PCA) to reduce the data dimensionality. To tackle the dataset imbalance issue, we re-weight our model loss by the inverse effective number of samples in each class. We use the human gait HuGaDB dataset with unique characteristics, for gait analysis.

Classification of weld defects using machine vision using convolutional neural network

Welding is an important aspect in commercial use of almost every industry. Because weld flaws can cause irregularities or inconsistencies during welding process, welding quality control is a critical step in ensuring the product’s quality and overall longevity. This study focuses on recognizing contamination defects, lack of fusion defects, or if the weld belongs to the good weld category among the defects that occur during the welding process. This category categorization is carried out for the Convolutional Neural Network (CNN) algorithm and the accuracy metric is obtained to evaluate the efficiency of the algorithm for the 3 – class dataset. According to this research, the pure CNN approach gave an accuracy result of 96.1%.

Short‐term and long‐term memory self‐attention network for segmentation of tumours in 3D medical images

AbstractTumour segmentation in medical images (especially 3D tumour segmentation) is highly challenging due to the possible similarity between tumours and adjacent tissues, occurrence of multiple tumours and variable tumour shapes and sizes. The popular deep learning‐based segmentation algorithms generally rely on the convolutional neural network (CNN) and Transformer. The former cannot extract the global image features effectively while the latter lacks the inductive bias and involves the complicated computation for 3D volume data. The existing hybrid CNN‐Transformer network can only provide the limited performance improvement or even poorer segmentation performance than the pure CNN. To address these issues, a short‐term and long‐term memory self‐attention network is proposed. Firstly, a distinctive self‐attention block uses the Transformer to explore the correlation among the region features at different levels extracted by the CNN. Then, the memory structure filters and combines the above information to exclude the similar regions and detect the multiple tumours. Finally, the multi‐layer reconstruction blocks will predict the tumour boundaries. Experimental results demonstrate that our method outperforms other methods in terms of subjective visual and quantitative evaluation. Compared with the most competitive method, the proposed method provides Dice (82.4% vs. 76.6%) and Hausdorff distance 95% (HD95) (10.66 vs. 11.54 mm) on the KiTS19 as well as Dice (80.2% vs. 78.4%) and HD95 (9.632 vs. 12.17 mm) on the LiTS.

Exploring the Applicability of Transfer Learning and Feature Engineering in Epilepsy Prediction Using Hybrid Transformer Model

Epilepsy prediction algorithms offer patients with drug-resistant epilepsy a way to reduce unintended harm from sudden seizures. The purpose of this study is to investigate the applicability of transfer learning (TL) technique and model inputs for different deep learning (DL) model structures, which may provide a reference for researchers to design algorithms. Moreover, we also attempt to provide a novel and precise Transformer-based algorithm. Two classical feature engineering methods and the proposed method which consists of various EEG rhythms are explored, then a hybrid Transformer model is designed to evaluate the advantages over pure convolutional neural networks (CNN)-based models. Finally, the performances of two model structures are analyzed utilizing patient-independent approach and two TL strategies. We tested our method on the CHB-MIT scalp EEG database, the results showed that our feature engineering method gains a significant improvement in model performance and is more suitable for Transformer-based model. In addition, the performance improvement of Transformer-based model utilizing fine-tuning strategies is more robust than that of pure CNN-based model, and our model achieved an optimal sensitivity of 91.7% with false positive rate (FPR) of 0.00/h. Our epilepsy prediction method achieves excellent performance and demonstrates its advantage over pure CNN-based structure in TL. Moreover, we find that the information contained in the gamma ( γ ) rhythm is helpful for epilepsy prediction. We propose a precise hybrid Transformer model for epilepsy prediction. The applicability of TL and model inputs is also explored for customizing personalized models in clinical application scenarios.

CNN hyper-parameter optimization for environmental sound classification

Environmental sounds are being used widely in our lives. It is especially used in tasks such as managing smart cities, location determination, surveillance systems, machine hearing, and environmental monitoring. The main method for this, environmental sound classification (ESC), has been increasingly studied in recent years. However, the classification of these sounds is more difficult than other sounds because there are too many parameters that generate noise. The study tried to find the convolutional neural network (CNN) model that gave the highest accuracy for ESC tasks with the optimization of hyper-parameters. For this purpose, the Particle Swarm Optimization (PSO) algorithm was rearranged to represent the CNN architecture. Thus, the hyper-parameters in CNN are represented exactly without any transformation during optimization. Studies were carried out on the ESC-10, ESC-50, and Urbansound8k data sets, which are state-of-art for ESC tasks. Some data augmentation techniques have been used for data sets in the training of CNN models. The CNN models, which were obtained with PSO, achieved success rates of 98.64 % for ESC-10, 93.71 % for ESC-50, and 98.45 % for Urbansound8k, respectively. These results are the best accuracy values obtained with the pure CNN model when compared with previous studies. As a result, it has been made possible to automatically design CNN models for the classification of urban sounds, giving high classification accuracy. Thus, researchers who do not know much about CNN design can use this method in their desired datasets without the need for expert knowledge.

Knowledge‐Informed Deep Neural Networks for Solar Flare Forecasting

AbstractRecently, although various deep learning techniques have been applied to building space weather prediction models, a large amount of relevant prior knowledge of solar eruptions and magnetic properties is ignored during the model development. By integrating prior knowledge in flare production into the convolutional neural network (CNN) structures, we have developed a knowledge‐informed deep neural network model aiming at forecasting solar flares. The line‐of‐sight magnetograms of Space‐weather HMI Active Region Patches (SHARP) from May 2010 to December 2018 are selected. We have surveyed the relationships between solar flares and both the active region (AR) area and magnetic type. When integrating prior knowledge into the CNNs, three aspects are considered: (a) keeping the magnetic structure unchanged (data preprocessing) while filling SHARP magnetograms into squares, (b) grouping the data samples into two subsets according to different flare productivities (sample grouping), and (c) adding AR area as an extra input parameter to the CNN (extra input parameter implementation). Pure CNN model, Fusion model 1, and Fusion model 2 are built to forecast M‐class or above flares in the next 48 hr, which involve data preprocessing, data preprocessing and sample grouping, and all the three aspects, respectively. Fusion model 2 that augments the most prior knowledge has the best performance. Our results imply that prior knowledge can play an important role in building deep learning flare forecasting models. In the future, adopting knowledge‐informed deep neural networks will be an effective way to further improve the forecasting performance for other space weather events.

ITUnet: Integration Of Transformers And Unet For Organs-At-Risk Segmentation.

Recently, convolutional neural network(CNN) has achieved great success in medical image segmentation. However, due to the limitation of convolutional receptive field, the pure convolutional neural network is difficult to further improve its performance. Given the outstanding ability of transformers in extracting the long-range dependency, some works have successfully applied it to computer vision and achieved better results than CNN in some tasks. Based on transformers could remedy the shortage of CNN, in this paper, we propose ITUnet, a segmentation network using CNN and transformers as features extractor. The combination of CNN and transformers enables the network to learn both short- and long-range dependency of features, which is beneficial to segmentation tasks. We evaluate our method on a head-and-neck CT dataset which has 18 kinds of organs to be segmented. The experimental results demonstrate that our proposed method shows better accuracy and robustness, the proposed methods achieve the Dice score of 77.72 and the 95% Hausdorff Distance of 2.31, outperforming the existing methods.

Topology-Aware Convolutional Neural Network for Efficient Skeleton-Based Action Recognition

In the context of skeleton-based action recognition, graph convolutional networks (GCNs) have been rapidly developed, whereas convolutional neural networks (CNNs) have received less attention. One reason is that CNNs are considered poor in modeling the irregular skeleton topology. To alleviate this limitation, we propose a pure CNN architecture named Topology-aware CNN (Ta-CNN) in this paper. In particular, we develop a novel cross-channel feature augmentation module, which is a combo of map-attend-group-map operations. By applying the module to the coordinate level and the joint level subsequently, the topology feature is effectively enhanced. Notably, we theoretically prove that graph convolution is a special case of normal convolution when the joint dimension is treated as channels. This confirms that the topology modeling power of GCNs can also be implemented by using a CNN. Moreover, we creatively design a SkeletonMix strategy which mixes two persons in a unique manner and further boosts the performance. Extensive experiments are conducted on four widely used datasets, i.e. N-UCLA, SBU, NTU RGB+D and NTU RGB+D 120 to verify the effectiveness of Ta-CNN. We surpass existing CNN-based methods significantly. Compared with leading GCN-based methods, we achieve comparable performance with much less complexity in terms of the required GFLOPs and parameters.

Deep learning-based prediction of traffic accidents risk for Internet of vehicles

With the increasing number of vehicles, traffic accidents pose a great threat to human lives. Hence, aiming at reducing the occurrence of traffic accidents, this paper proposes an algorithm based on a deep convolutional neural network and a random forest to predict accident risks. Specifically, the proposed algorithm includes a feature extractor and a feature classifier, where the former extracts key features using a convolutional neural network and the latter outputs a probability value of traffic accidents using a random forest with multiple decision trees, which indicates the degree of accident risks. Simulations show that the proposed algorithm can achieve higher performance in terms of the Area Under the Curve (AUC) of the Receiver Characteristic Operator as well as accuracy than the existing algorithms based on the Adaboost or the pure convolutional neural networks.