Output Feature Maps Research Articles

Significant feature suppression and cross-feature fusion networks for fine-grained visual classification

The technique of extracting different distinguishing features by locating different part regions to achieve fine-grained visual classification (FGVC) has made significant improvements. Utilizing attention mechanisms for feature extraction has become one of the mainstream methods in computer vision, but these methods have certain limitations. They typically focus on the most discriminative regions and directly combine the features of these parts, neglecting other less prominent yet still discriminative regions. Additionally, these methods may not fully explore the intrinsic connections between higher-order and lower-order features to optimize model classification performance. By considering the potential relationships between different higher-order feature representations in the object image, we can enable the integrated higher-order features to contribute more significantly to the model’s classification decision-making capabilities. To this end, we propose a saliency feature suppression and cross-feature fusion network model (SFSCF-Net) to explore the interaction learning between different higher-order feature representations. These include (1) an object-level image generator (OIG): the intersection of the output feature maps of the last two convolutional blocks of the backbone network is used as an object mask and mapped to the original image for cropping to obtain an object-level image, which can effectively reduce the interference caused by complex backgrounds. (2) A saliency feature suppression module (SFSM): the most distinguishing part of the object image is obtained by a feature extractor, and the part is masked by a two-dimensional suppression method, which improves the accuracy of feature suppression. (3) A cross-feature fusion method (CFM) based on inter-layer interaction: the output feature maps of different network layers are interactively integrated to obtain high-dimensional features, and then the high-dimensional features are channel compressed to obtain the inter-layer interaction feature representation, which enriches the output feature semantic information. The proposed SFSCF-Net can be trained end-to-end and achieves state-of-the-art or competitive results on four FGVC benchmark datasets.

MobileNet-V2: An Enhanced Skin Disease Classification by Attention and Multi-Scale Features.

The increasing prevalence of skin diseases necessitates accurate and efficient diagnostic tools. This research introduces a novel skin disease classification model leveraging advanced deep learning techniques. The proposed architecture combines the MobileNet-V2 backbone, Squeeze-and-Excitation (SE) blocks, Atrous Spatial Pyramid Pooling (ASPP), and a Channel Attention Mechanism. The model was trained on four diverse datasets such as PH2 dataset, Skin Cancer MNIST: HAM10000 dataset, DermNet. dataset, and Skin Cancer ISIC dataset. Data preprocessing techniques, including image resizing, and normalization, played a crucial role in optimizing model performance. In this paper, the MobileNet-V2 backbone is implemented to extract hierarchical features from the preprocessed dermoscopic images. The multi-scale contextual information is fused by the ASPP model for generating a feature map. The attention mechanisms contributed significantly, enhancing the extraction ability of inter-channel relationships and multi-scale contextual information for enhancing the discriminative power of the features. Finally, the output feature map is converted into probability distribution through the softmax function. The proposed model outperformed several baseline models, including traditional machine learning approaches, emphasizing its superiority in skin disease classification with 98.6% overall accuracy. Its competitive performance with state-of-the-art methods positions it as a valuable tool for assisting dermatologists in early classification. The study also identified limitations and suggested avenues for future research, emphasizing the model's potential for practical implementation in the field of dermatology.

Multiple myeloma segmentation net (MMNet): an encoder-decoder-based deep multiscale feature fusion model for multiple myeloma segmentation in magnetic resonance imaging.

Patients with multiple myeloma (MM), a malignant disease involving bone marrow plasma cells, shows significant susceptibility to bone degradation, impairing normal hematopoietic function. The accurate and effective segmentation of MM lesion areas is crucial for the early detection and diagnosis of myeloma. However, the presence of complex shape variations, boundary ambiguities, and multiscale lesion areas, ranging from punctate lesions to extensive bone damage, presents a formidable challenge in achieving precise segmentation. This study thus aimed to develop a more accurate and robust segmentation method for MM lesions by extracting rich multiscale features. In this paper, we propose a novel, multiscale feature fusion encoding-decoding model architecture specifically designed for MM segmentation. In the encoding stage, our proposed multiscale feature extraction module, dilated dense connected net (DCNet), is employed to systematically extract multiscale features, thereby augmenting the model's sensing field. In the decoding stage, we propose the CBAM-atrous spatial pyramid pooling (CASPP) module to enhance the extraction of multiscale features, enabling the model to dynamically prioritize both channel and spatial information. Subsequently, these features are concatenated with the final output feature map to optimize segmentation outcomes. At the feature fusion bottleneck layer, we incorporate the dynamic feature fusion (DyCat) module into the skip connection to dynamically adjust feature extraction parameters and fusion processes. We assessed the efficacy of our approach using a proprietary dataset of MM, yielding notable advancements. Our dataset comprised 753 magnetic resonance imaging (MRI) two-dimensional (2D) slice images of the spinal regions from 45 patients with MM, along with their corresponding ground truth labels. These images were primarily obtained from three sequences: T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and short tau inversion recovery (STIR). Using image augmentation techniques, we expanded the dataset to 3,000 images, which were employed for both model training and prediction. Among these, 2,400 images were allocated for training purposes, while 600 images were reserved for validation and testing. Our method showed increase in the intersection over union (IoU) and Dice coefficients by 7.9 and 6.7 percentage points, respectively, as compared to the baseline model. Furthermore, we performed comparisons with alternative image segmentation methodologies, which confirmed the sophistication and efficacy of our proposed model. Our proposed multiple myeloma segmentation net (MMNet), can effectively extract multiscale features from images and enhance the correlation between channel and spatial information. Furthermore, a systematic evaluation of the proposed network architecture was conducted on a self-constructed, limited dataset. This endeavor holds promise for offering valuable insights into the development of algorithms for future clinical applications.

FFA-BiGRU: Attention-Based Spatial-Temporal Feature Extraction Model for Music Emotion Classification

Music emotion recognition is becoming an important research direction due to its great significance for music information retrieval, music recommendation, and so on. In the task of music emotion recognition, the key to achieving accurate emotion recognition lies in how to extract the affect-salient features fully. In this paper, we propose an end-to-end spatial-temporal feature extraction method named FFA-BiGRU for music emotion classification. Taking the log Mel-spectrogram of music audio as the input, this method employs an attention-based convolutional residual module named FFA, which serves as a spatial feature learning module to obtain multi-scale spatial features. In the FFA module, three group architecture blocks extract multi-level spatial features, each of which consists of a stack of multiple channel-spatial attention-based residual blocks. Then, the output features from FFA are fed into the bidirectional gated recurrent units (BiGRU) module to capture the temporal features of music further. In order to make full use of the extracted spatial and temporal features, the output feature maps of FFA and those of the BiGRU are concatenated in the channel dimension. Finally, the concatenated features are passed through fully connected layers to predict the emotion classification results. The experimental results of the EMOPIA dataset show that the proposed model achieves better classification accuracy than the existing baselines. Meanwhile, the ablation experiments also demonstrate the effectiveness of each part of the proposed method.

FPWT: Filter pruning via wavelet transform for CNNs

The enormous data and computational resources required by Convolutional Neural Networks (CNNs) hinder the practical application on mobile devices. To solve this restrictive problem, filter pruning has become one of the practical approaches. At present, most existing pruning methods are currently developed and practiced with respect to the spatial domain, which ignores the potential interconnections in the model structure and the decentralized distribution of image energy in the spatial domain. The image frequency domain transform method can remove the correlation between image pixels and concentrate the image energy distribution, which results in lossy compression of images. In this study, we find that the frequency domain transform method is also applicable to the feature maps of CNNs. The filter pruning via wavelet transform (WT) is proposed in this paper (FPWT), which combines the frequency domain information of WT with the output feature map to more obviously find the correlation between feature maps and make the energy into a relatively concentrated distribution in the frequency domain. Moreover, the importance score of each feature map is calculated by the cosine similarity and the energy-weighted coefficients of the high and low frequency components, and prune the filter based on its importance score. Experiments on two image classification datasets validate the effectiveness of FPWT. For ResNet-110 on CIFAR-10, FPWT reduces FLOPs and parameters by more than 60.0 % with 0.53 % accuracy improvement. For ResNet-50 on ImageNet, FPWT reduces FLOPs by 53.8 % and removes parameters by 49.7 % with only 0.97 % loss of Top-1 accuracy.

An Identification Method for Mixed Coal Vitrinite Components Based on An Improved DeepLabv3+ Network

To address the high complexity and low accuracy issues of traditional methods in mixed coal vitrinite identification, this paper proposes a method based on an improved DeepLabv3+ network. First, MobileNetV2 is used as the backbone network to reduce the number of parameters. Second, an atrous convolution layer with a dilation rate of 24 is added to the ASPP (atrous spatial pyramid pooling) module to further increase the receptive field. Meanwhile, a CBAM (convolutional block attention module) attention mechanism with a channel multiplier of 8 is introduced at the output part of the ASPP module to better filter out important semantic features. Then, a corrective convolution module is added to the network’s output to ensure the consistency of each channel’s output feature map for each type of vitrinite. Finally, images of 14 single vitrinite components are used as training samples for network training, and a validation set is used for identification testing. The results show that the improved DeepLabv3+ achieves 6.14% and 3.68% improvements in MIOU (mean intersection over union) and MPA (mean pixel accuracy), respectively, compared to the original DeepLabv3+; 12% and 5.3% improvements compared to U-Net; 9.26% and 4.73% improvements compared to PSPNet with ResNet as the backbone; 5.4% and 9.34% improvements compared to PSPNet with MobileNetV2 as the backbone; and 6.46% and 9.05% improvements compared to HRNet. Additionally, the improved ASPP module increases MIOU and MPA by 3.23% and 1.93%, respectively, compared to the original module. The CBAM attention mechanism with a channel multiplier of 8 improves MIOU and MPA by 1.97% and 1.72%, respectively, compared to the original channel multiplier of 16. The data indicate that the proposed identification method significantly improves recognition accuracy and can be effectively applied to mixed coal vitrinite identification.

Defect detection of photovoltaic modules based on improved VarifocalNet

Detecting and replacing defective photovoltaic modules is essential as they directly impact power generation efficiency. Many current deep learning-based methods for detecting defects in photovoltaic modules focus solely on either detection speed or accuracy, which limits their practical application. To address this issue, an improved VarifocalNet has been proposed to enhance both the detection speed and accuracy of defective photovoltaic modules. Firstly, a new bottleneck module is designed to replace the first bottleneck module of the last stage convolution group in the backbone. This new module includes both standard convolution and dilated convolution, enabling an increase in network depth and receptive field without reducing the output feature map size. This improvement can help to enhance the accuracy of defect detection for photovoltaic modules. Secondly, another bottleneck module is also designed and used to replace the original bottleneck module used in the fourth stage convolution group of the backbone. This new module has smaller parameters than the original bottleneck module, which is useful to improve the defect detection speed of the photovoltaic module. Thirdly, a feature interactor is designed in the detection head to enhance feature expression in the classification branch. This helps improve detection accuracy. Besides, an improved intersection over union is proposed and introduced into the loss function to measure the difference between the predicted and ground truth boxes. This is useful for improving defect detection accuracy. Compared to other methods, the proposed method has the highest detection accuracy. Additionally, it also has a faster detection speed than other methods except for the DDH-YOLOv5 method and the improved YOLOv7 method.

Efficient filter pruning: Reducing model complexity through redundancy graph decomposition

Filter pruning has emerged as a crucial technique facilitating the efficient deployment of Convolutional Neural Networks (CNNs) on edge devices. The resultant lightweight models often present a challenging trade-off between performance and compression rate. Moreover, the excessive granularity of filter selection, coupled with intricate fine-tuning procedures, hampers the practical applicability of numerous pruning methodologies. To tackle these limitations, our paper introduces a novel pruning approach that initially constructs an undirected graph, effectively capturing convolutional layer redundancy by quantifying the similarity among output feature maps. Subsequently, our method iteratively eliminates highly redundant and low-value filters, employing both k-core decomposition and importance ranking criteria. In practical implementation, we adopt a one-shot pruning strategy, thereby minimizing the requirement for extensive fine-tuning. Comprehensive comparative experiments demonstrate the efficacy of our proposed approach in achieving remarkable improvements in model compression without compromising performance. For instance, with CIFAR-10, our method achieves state-of-the-art results on DenseNet-40 with a mere 70 fine-tuning epochs, while effectively removing 49.39% of FLOPs, resulting in a mere 0.01% accuracy loss. Furthermore, in the context of ImageNet, our approach excels by removing 48.81% of FLOPs and 23.71% of parameters from ResNet-50, exhibiting only 1.47% and 0.83% decline in Top-1 and Top-5 Accuracy respectively. The code is publicly available at https://github.com/lijiang99/redundancy-graph-decomposition.

Weakly supervised instance segmentation via class double-activation maps and boundary localization

Weakly supervised instance segmentation based on image-level class labels has recently gained much attention, in which the primary key step is to generate the pseudo labels based on class activation maps (CAMs). Most methods adopt binary cross-entropy (BCE) loss to train the classification model. However, since BCE loss is not class mutually exclusive, activations among classes occur independently. Thus, not only do foreground classes are wrongly activated as background, but also incorrect activations among confusing classes are occurred in the foreground. To solve this problem, we propose the Class Double-Activation Map, called Double-CAM. Firstly, the vanilla CAM is extracted from the multi-label classifier and then fused with the output feature map of backbone. The enhanced feature map of each class is fed into the single-label classification branch with softmax cross-entropy (SCE) loss and entropy minimization module, from which the more accurate Double-CAM is extracted. It refines the vanilla CAM to improve the quality of pseudo labels. Secondly, to mine object edge cues from Double-CAM, we propose the Boundary Localization (BL) module to synthesize boundary annotations, so as to provide constraints for label propagation more explicitly without adding additional supervision. The quality of pseudo masks is also improved substantially with the addition of BL module. Finally, the generated pseudo labels are used to train fully supervised instance segmentation networks. The evaluations on VOC and COCO datasets show that our method achieves excellent performance, outperforming mainstream weakly supervised segmentation methods at the same supervisory level, even those that depend on stronger supervision.

Heterogeneous heatmap distillation framework based on unbiased alignment for lightweight human pose estimation

The growing demand for mobile devices has generated interest in lightweight human pose estimation. Currently, lightweight estimation generally uses heatmap-based methods, which has demonstrated exceptional performance. However, their use of non-differentiable post-processing imposes considerable inference latencies. Conversely, integral-based approaches expedite the inference process by employing a soft-argmax operation but compromise in accuracy. Integrating explicit heatmap knowledge learned using the heatmap-based method into the implicit heatmap generated by the integral-based method, thereby combining the best of both worlds, offers a promising avenue. However, owing to the disparities in supervision and inference processes, the explicit and implicit heatmaps are heterogeneous. Consequently, direct transfer of knowledge presents difficulties in ensuring consistencies in heat value and location. In this paper, we propose a novel Heterogeneous Heatmap Distillation (HHD) framework that effectively tackles these challenges. The framework seamlessly integrates explicit heatmap knowledge that contains high-precision localization information into implicit heatmaps. The framework revolves around an unbiased heatmap alignment scheme encompassing two steps: heterogeneous heatmap normalization and unbiased cropping. Heterogeneous heatmap normalization separately normalizes the output feature maps of both the teacher and student models, alleviating potential heat value bias during the knowledge transfer. Unbiased cropping applies closed-form computation on the normalized teacher and student heatmap to eliminate location bias. Additionally, mirror expansion is implemented to handle potential cases wherein the cropped region extends beyond the image boundary. Extensive experiments demonstrate the efficiency and effectiveness of our methods on the MSCOCO and MPII datasets compared to other integral-based lightweight networks. Our source codes and pre-trained models are available at https://github.com/ducongju/HHD.

Fluctuation-Based Adaptive Structured Pruning for Large Language Models

Network Pruning is a promising way to address the huge computing resource demands of the deployment and inference of Large Language Models (LLMs). Retraining-free is important for LLMs' pruning methods. However, almost all of the existing retraining-free pruning approaches for LLMs focus on unstructured pruning, which requires specific hardware support for acceleration. In this paper, we propose a novel retraining-free structured pruning framework for LLMs, named FLAP (FLuctuation-based Adaptive Structured Pruning). It is hardware-friendly by effectively reducing storage and enhancing inference speed. For effective structured pruning of LLMs, we highlight three critical elements that demand the utmost attention: formulating structured importance metrics, adaptively searching the global compressed model, and implementing compensation mechanisms to mitigate performance loss. First, FLAP determines whether the output feature map is easily recoverable when a column of weight is removed, based on the fluctuation pruning metric. Then it standardizes the importance scores to adaptively determine the global compressed model structure. At last, FLAP adds additional bias terms to recover the output feature maps using the baseline values. We thoroughly evaluate our approach on a variety of language benchmarks. Without any retraining, our method significantly outperforms the state-of-the-art methods, including LLM-Pruner and the extension of Wanda in structured pruning. The code is released at https://github.com/CASIA-IVA-Lab/FLAP.

TTDCapsNet: Tri Texton-Dense Capsule Network for complex and medical image recognition.

Convolutional Neural Networks (CNNs) are frequently used algorithms because of their propensity to learn relevant and hierarchical features through their feature extraction technique. However, the availability of enormous volumes of data in various variations is crucial for their performance. Capsule networks (CapsNets) perform well on a small amount of data but perform poorly on complex images. To address this, we proposed a new Capsule Network architecture called Tri Texton-Dense CapsNet (TTDCapsNet) for better complex and medical image classification. The TTDCapsNet is made up of three hierarchic blocks of Texton-Dense CapsNet (TDCapsNet) models. A single TDCapsNet is a CapsNet architecture composed of a texton detection layer to extract essential features, which are passed onto an eight-layered block of dense convolution that further extracts features, and then the output feature map is given as input to a Primary Capsule (PC), and then to a Class Capsule (CC) layer for classification. The resulting feature map from the first PC serves as input into the second-level TDCapsNet, and that from the second PC serves as input into the third-level TDCapsNet. The routing algorithm receives feature maps from each PC for the various CCs. Routing the concatenation of the three PCs creates an additional CC layer. All these four feature maps combined, help to achieve better classification. On fashion-MNIST, CIFAR-10, Breast Cancer, and Brain Tumor datasets, the proposed model is evaluated and achieved validation accuracies of 94.90%, 89.09%, 95.01%, and 97.71% respectively. Findings from this work indicate that TTDCapsNet outperforms the baseline and performs comparatively well with the state-of-the-art CapsNet models using different performance metrics. This work clarifies the viability of using Capsule Network on complex tasks in the real world. Thus, the proposed model can be used as an intelligent system, to help oncologists in diagnosing cancerous diseases and administering treatment required.

Hyperspectral image classification method based on narrowing semantic gap convolutional neural network

ABSTRACT In recent years, numerous hyperspectral image classification methods based on convolutional neural network (CNN) have been proposed in recent years. However, the majority of CNN-based classification models fail to take into account the semantic gap between shallow and deep features, and instead, just sum up or splice these two types of features, which leads to subpar feature fusion. To address the above problems, the paper proposes a hyperspectral image classification method based on Narrowing Semantic Gap Convolutional Neural Network (NSGCNN). We developed the feature fusion sub-network, which consists of Feature Semantic Enhancement Module (FSEM), Spatial Detail Supplement Module (SDCM), and Cross-layer Feature Fusion Module (CFFM). FSEM extracts rich global semantic information based on multi-scale dilated convolution. Reverse attention mechanism is introduced in the SDCM to complement deep-layer features with fine spatial features like edges and textures. The fusing of hierarchical features is then achieved using CFFM. Additionally, we can regulate the model’s parameter count by regulating the number of output feature maps. Experimental results on three benchmark hyperspectral datasets demonstrate that NSGCNN outperforms other state-of-the-art methods in terms of classification performance and significantly outperforms them in terms of model complexity.

A novel data-driven integrated detection method for network intrusion classification based on multi-feature imbalanced data

The multi-feature and imbalanced nature of network data has always been a challenge to be overcome in the field of network intrusion detection. The redundant features in data could reduce the overall quality of network data and the accuracy of detection models, because imbalance could lead to a decrease in the detection rate for minority classes. To improve the detection accuracy for imbalanced intrusion data, we develop a data-driven integrated detection method, which utilizes Recursive Feature Elimination (RFE) for feature selection, and screens out features that are conducive to model recognition for improving the overall quality of data analysis. In this work, we also apply the Adaptive Synthetic Sampling (ADASYN) method to generate the input data close to the original dataset, which aims to eliminate the data imbalance in the studied intrusion detection model. Besides, a novel VGG-ResNet classification algorithm is also proposed via integrating the convolutional block with the output feature map size of 128 from the Visual Geometry Group 16 (VGG16) of the deep learning algorithm and the residual block with output feature map size of 256 from the Residual Network 18 (ResNet18). Based on the numerical results conducted on the well-known NSL-KDD dataset and UNSW-NB15 dataset, it illustrates that our method can achieve the accuracy rates of 86.31% and 82.56% in those two test datasets, respectively. Moreover, it can be found that the present algorithm can achieve a better accuracy and performance in the experiments of comparing our method with several existing algorithms proposed in the recent three years.

ASAFormer: Visual tracking with convolutional vision transformer and asymmetric selective attention

Recently, Vision Transformer (ViT) has exhibited remarkable performances in many computer vision tasks (e.g. object detection, segmentation and tracking). However, the output feature map of ViT is only single scale with low resolution, which may lose rich detailed semantic information. Meanwhile, ViT implements feature embedding through the linear projection, which makes it unable to capture local spatial context. Furthermore, as the core component of Transformer, self-attention captures the long-range dependencies at the cost of large memory footprint during training. In this paper, a novel hierarchical model is proposed to remedy the above issues. Firstly, convolutional vision Transformer is employed as our backbone for feature extraction and fusion. Secondly, a novel asymmetric structure is presented to calculate the cross-relation of the template and search branches. Thirdly, different selection operations are devised for the input of the attention module in both branches. Extensive experiments have been conducted on 5 mainstream benchmarks, which exhibits the superiority of our tracker. The code will be available at here.

TAILOR: InTer-feAture distinctIon fiLter fusiOn pRuning

Filter pruning is an effective method for reducing the size of convolutional neural networks without sacrificing performance. Most filter pruning methods prioritize filters with high information content, but fail to consider that filters with low information content might capture essential features. Moreover, we have discovered that the distinctions among feature maps generated by filters can identify crucial features. Based on this insight, we propose a novel pruning method called inTer-feAture distinctIon fiLter fusiOn pRuning (TAILOR), which fuses the feature distinctions between filters. TAILOR randomly selects multiple filter sets within a convolutional layer and calculates the output feature maps of the next convolutional layer of these sets. Subsequently, an intelligent distinction optimization scheme is proposed to obtain the optimal filter set for filter pruning, which supplants the original convolutional layer. Experimental results indicated that the inter-feature distinctions among filters significantly affect filter pruning. TAILOR outperforms state-of-the-art filter-pruning methods in terms of model prediction accuracy, floating-point operations, and parameter scale. For instance, with VGG-16, TAILOR achieves a 73.89% FLOPs reduction by removing 91.85% of the parameters, while improving accuracy by 0.36% on the CIFAR-10.

FCPFNet: Feature Complementation Network with Pyramid Fusion for Semantic Segmentation

Traditional pyramid pooling modules have shown effective improvements in semantic segmentation tasks by capturing multi-scale feature information. However, their limitations arise from the shallow structure, which fails to fully extract contextual information, and the fused multi-scale feature information lacks distinctiveness, resulting in issues with the final segmentation discriminability. To address these issues, we proposes an effective solution called FCPFNet, which is based on global contextual prior for deep feature extraction of detailed information. Specifically, we introduce a novel deep feature aggregation module to extract semantic information from the output feature map of each layer through a deep aggregation of context information module, and expands the effective perception range. Additionally, we propose an Efficient Pyramid Pooling Module (EPPM) to capture distinctive features through communicating information between different sub-features and performs multi-scale fusion, which is integrated as a branch within the network to complement the information loss resulting from downsampling operations. Furthermore, in order to ensure the richness of image detail feature information and maintain a large receptive field to obtain more contextual information, EPPM concatenates the input feature map and the output feature map of the pyramid pooling module to acquire more comprehensive global contextual information. It has been demonstrated by experiment that the method described in this article achieves competitive performance on the challenging scene segmentation datasets Pascal VOC 2012, Cityscapes and Coco-Stuff, with MIOU of 81.0%, 78.8% and 40.1%, respectively.

Enhancing artistic analysis through deep learning: a graphic art element recognition model based on SSD and FPT.

For the analysis of art works, accurate identification of various elements of works through deep learning methods is helpful for artists to appreciate and learn works. In this study, we leverage deep learning methodologies to precisely identify the diverse elements within graphic art designs, aiding artists in their appreciation and learning process. Our approach involves integrating the attention mechanism into an enhanced Single Shot MultiBox Detector (SSD) model to refine the recognition of artistic design elements. Additionally, we improve the feature fusion structure of the SSD model by incorporating long-range attention mechanism information, thus enhancing target detection accuracy. Moreover, we refine the Feature Pyramid Transformer (FPT) attention mechanism model to ensure the output feature map aligns effectively with the requirements of object detection. Our empirical findings demonstrate that our refined approach outperforms the original SSD algorithm across all four evaluation metrics, exhibiting improvements of 1.52%, 1.89%, 3.09%, and 2.57%, respectively. Qualitative tests further illustrate the accuracy, robustness, and universality of our proposed method, particularly in scenarios characterized by dense artistic elements and challenging-to-distinguish categories within art compositions.

Class-Separation Preserving Pruning for Deep Neural Networks

Neural network pruning has been deemed essential in the deployment of deep neural networks on resource-constrained edge devices, greatly reducing the number of network parameters without drastically compromising accuracy. A class of techniques proposed in the literature assigns an importance score to each parameter and prunes those of the least importance. However, most of these methods are based on generalised estimations of the importance of each parameter, ignoring the context of the specific task at hand. In this paper, we propose a task specific pruning approach, CSPrune, which is based on how efficiently a neuron or a convolutional filter is able to separate classes. Our axiomatic approach assigns an importance score based on how separable different classes are in the output activations or feature maps, preserving separation of classes which avoids reduction in classification accuracy. Additionally, most pruning algorithms prune individual connections or weights leading to a sparse network without taking into account whether the hardware the network is deployed on can take advantage of that sparsity or not. CSPrune prunes whole neurons or filters which results in a more structured pruned network whose sparsity can be more efficiently utilised by the hardware. We evaluate our pruning method against various benchmark datasets, both small and large, and network architectures and show that our approach outperforms comparable pruning techniques.

Sparse CNN and Deep Reinforcement Learning-Based D2D Scheduling in UAV-Assisted Industrial IoT Networks

Unmanned aerial vehicles (UAVs) have been widely applied in wireless communications because of its high flexibility and line-of-sight transmission. In this study, we develop low-complexity and robust device-to-device (D2D) link scheduling in UAV-assisted industrial Internet of Things (IIoT) networks. First, we propose a sparse convolutional neural network (SCNN) model that uses the geographical map of transmission links as input. The model consists of three main blocks: generic feature filtering, speed–accuracy balancing, and deep feature processing. Unlike other state-of-the-art methods, the proposed SCNN directly processes the geographical map collected using a connected UAV. Second, we propose a deep deterministic policy gradient-based reinforcement learning model that processes the output feature map from the SCNN to optimize the D2D scheduling decision and maximize the achievable system rate in the long run. Extensive simulations revealed that the proposed scheme significantly improved the achievable rate over other benchmark comparison schemes, such as transmitters and receivers (T&R) density-based deep learning (DL), ResNet-based DL, VGGNet-based DL, random scheduling, and all-active schemes, respectively. The simulations also demonstrated that the proposed scheme reduces computational complexity. With reduced complexity and nearly optimal performance, the proposed solution can be more efficiently applied to large-scale and dense IIoT networks.