Improvement In Average Precision Research Articles

Lightweight DB-YOLO Facemask Intelligent Detection and Android Application Based on Bidirectional Weighted Feature Fusion

Conventional facemask detection algorithms face challenges of insufficient accuracy, large model size, and slow computation speed, limiting their deployment in real-world scenarios, especially on edge devices. Aiming at addressing these issues, we proposed a DB-YOLO facemask intelligent detection algorithm, which is a lightweight solution that leverages bidirectional weighted feature fusion. Our method is built on the YOLOv5 algorithm model, replacing the original YOLOv5 backbone network with the lightweight ShuffleNetv2 to reduce parameters and computational requirements. Additionally, we integrated BiFPN as the feature fusion layer, enhancing the model’s detection capability for objects of various scales. Furthermore, we employed a CARAFE lightweight upsampling factor to improve the model’s perception of details and small-sized objects and the EIOU loss function to expedite model convergence. We validated the effectiveness of our proposed method through experiments conducted on the Pascal VOC2007+2012 and Face_Mask datasets. Our experimental results demonstrate that the DB-YOLO model boasts a compact size of approximately 1.92 M. It achieves average precision values of 70.1% and 93.5% on the Pascal VOC2007+2012 and Face_Mask datasets, respectively, showcasing a 2.3% improvement in average precision compared to the original YOLOv5s. Furthermore, the model’s size is reduced by 85.8%. We also successfully deployed the model on Android devices using the NCNN framework, achieving a detection speed of up to 33 frames per second. Compared to lightweight algorithm models like YOLOv5n, YOLOv4-Tiny, and YOLOv3-Tiny, DB-YOLO not only reduces the model’s size but also effectively improves detection accuracy, exhibiting excellent practicality and promotional value on edge devices.

A high-accuracy and lightweight detector based on a graph convolution network for strip surface defect detection

For strip surface defect detection, the key is to achieve reliable detection results with high detection speed. This paper mainly focuses on the ability to distinguish defects with similar optical characteristics, and the balance between detection accuracy and speed. Firstly, the dataset with 2020 pictures containing 6 types of defects was established by the figures inspected in a rolled titanium strip production line. Then, a novel detection model named Yolo-SAGC was proposed by applying two strategies to the fast response Yolo-v5 model. One is to improve the feature recognition capability by combining self-attention and graphic convolution in the head module. The other is to make a thorough slim of the whole network architecture by using slim modules combined with depth-wise separable convolution (i.e., DWconv). Finally, the advancement of this novel detection model was verified by the self-established database. The results demonstrate a significant reduction in cases where detection is missed for the 6 types of defects, dropping from 32.75% to 6.67% when the two strategies are implemented. Notably, the most difficult-to-detect label “Pit” defect shows an 11.9% improvement in average precision with the introduction of self-attentional graphic convolution. Similarly, the densely distributed small target “Little_lightspot” exhibits a 5.0% increase in average precision when DWconv is applied. Furthermore, the mAP@0.5 of Yolo-SAGC is comparable to that of Yolo-v8, while the model parameters are decreased by 48.7% and FPS is increased by 3. These phenomena show the great potential of Yolo-SAGC in industrial applications.

Small object detection in unmanned aerial vehicle images using multi-scale hybrid attention

Small object detection in unmanned aerial vehicle images is always challenging due to the low resolution and the limited amount of information that they contain. Many feature enhancement effects have been introduced to improve the detection of small objects, but the extracted effective information is still insufficient, and redundant information interference is an issue. In this paper, we propose a new multi-scale hybrid attention based detector (MHA-YOLOv5), which integrates the similarity relationships between objects into you only look once version 5 (YOLOv5) for small object detection. Specifically, a novel multi-scale hybrid attention (MHA) structure is proposed to enhance the feature representation of small objects. This structure contains three modules: multi-scale attention (MsA), foreground enhancement module (FEM) and depthwise separable channel attention (DSCA). The MsA module is designed to build connections between large objects with abundant details and small objects with insufficient features on multiple scale features and capture the similarity relationships between objects. To reduce the interference of redundant information, the FEM is used to focus on the foreground features of multiple scale features, and the DSCA module is utilized to effectively extract multidimensional channel information. Sufficient experiments on the challenging VisDrone2019-DET, UAVDT and CARPK datasets demonstrate the effectiveness and superiority of the proposed approach. Specifically, compared with the performance of YOLOv5, MHA-YOLOv5 demonstrates a 2.82% mean average precision (mAP) improvement on the VisDrone2019-DET dataset, a 2.25% mAP improvement on the UAVDT dataset, and a 3.07% mAP improvement on the CARPK dataset.

MT-Det: A novel fast object detector of maize tassel from high-resolution imagery using single level feature

Accurately and quickly detecting maize tassels is crucial for successful maize breeding and seed production. However, the complex planting environment in the field, including unsynchronized growth stages, cluttered backgrounds, severe occlusions, and varying tassel sizes and shapes, presents a challenge for detection methods. Current methods either use mainstream object detectors off-the-shelf or employ shallow adaptations on the modular level, lacking a systematic design that balances both speed and accuracy. Furthermore, most studies only report evaluation results on sliced image patches instead of original high-resolution images, which can offer limited guidance for large-scale crop management. In this study, we propose a novel one-stage anchor-free maize tassel detector (MT-Det), which is based on a single level feature and designed to be simple yet effective. Extensive comparisons with both one-level counterparts and feature pyramid detectors demonstrate that MT-Det outperforms them in both detection accuracy and inference speed. To address the issue of significant accuracy drop when directly inferring on high-resolution images, we introduce a bundle of slicing-aided hyper inference technologies to our proposed MT-Det, which results in a 13% and 38% improvement of mean average precision (mAP) on proximal and unmanned aerial vehicle (UAV) high-resolution images, respectively. We believe that MT-Det provides a promising high-throughput solution for accurate and efficient detection and counting of maize tassels in real-world field conditions.

Small-Sample Underwater Target Detection: A Joint Approach Utilizing Diffusion and YOLOv7 Model

Underwater target detection technology plays a crucial role in the autonomous exploration of underwater vehicles. In recent years, significant progress has been made in the field of target detection through the application of artificial intelligence technology. Effectively applying AI techniques to underwater target detection is a highly promising area of research. However, the difficulty and high cost of underwater acoustic data collection have led to a severe lack of data, greatly restricting the development of deep-learning-based target detection methods. The present study is the first to utilize diffusion models for generating underwater acoustic data, thereby effectively addressing the issue of poor detection performance arising from the scarcity of underwater acoustic data. Firstly, we place iron cylinders and cones underwater (simulating small preset targets such as mines). Subsequently, we employ an autonomous underwater vehicle (AUV) equipped with side-scan sonar (SSS) to obtain underwater target data. The collected target data are augmented using the denoising diffusion probabilistic model (DDPM). Finally, the augmented data are used to train an improved YOLOv7 model, and its detection performance is evaluated on a test set. The results demonstrate the effectiveness of the proposed method in generating similar data and overcoming the challenge of limited training sample data. Compared to models trained solely on the original data, the model trained with augmented data shows a mean average precision (mAP) improvement of approximately 30% across various mainstream detection networks. Additionally, compared to the original model, the improved YOLOv7 model proposed in this study exhibits a 2% increase in mAP on the underwater dataset.

YOLO-SS-Large: A Lightweight and High-Performance Model for Defect Detection in Substations.

With the development of deep fusion intelligent control technology and the application of low-carbon energy, the number of renewable energy sources connected to the distribution grid has been increasing year by year, gradually replacing traditional distribution grids with active distribution grids. In addition, as an important component of the distribution grid, substations have a complex internal environment and numerous devices. The problems of untimely defect detection and slow response during intelligent inspections are particularly prominent, posing risks and challenges to the safe and stable operation of active distribution grids. To address these issues, this paper proposes a high-performance and lightweight substation defect detection model called YOLO-Substation-large (YOLO-SS-large) based on YOLOv5m. The model improves lightweight performance based upon the FasterNet network structure and obtains the F-YOLOv5m model. Furthermore, in order to enhance the detection performance of the model for small object defects in substations, the normalized Wasserstein distance (NWD) and complete intersection over union (CIoU) loss functions are weighted and fused to design a novel loss function called NWD-CIoU. Lastly, based on the improved model mentioned above, the dynamic head module is introduced to unify the scale-aware, spatial-aware, and task-aware attention of the object detection heads of the model. Compared to the YOLOv5m model, the YOLO-SS-Large model achieves an average precision improvement of 0.3%, FPS enhancement of 43.5%, and parameter reduction of 41.0%. This improved model demonstrates significantly enhanced comprehensive performance, better meeting the requirements of the speed and precision for substation defect detection, and plays an important role in promoting the informatization and intelligent construction of active distribution grids.

RodNet: An Advanced Multidomain Object Detection Approach Using Feature Transformation With Generative Adversarial Networks

Advanced object detection techniques have been widely studied in recent years and have been successfully applied in real-world applications. However, existing algorithms may struggle with nighttime image detection, especially in low-luminance conditions. Researchers have attempted to overcome this issue by collecting large amounts of multi-domain data, but performance remains poor because these methods train images from both low-luminance and sufficient-luminance domains without a specific training policy. In this work, we present a lightweight framework for multi-domain object detection using feature domain transformation with generative adversarial networks (GANs). The proposed GAN framework trains a generator network to transform features from the low-luminance domain to a sufficient-luminance domain, making the discriminator networks unable to distinguish whether the features were generated from a low-luminance or a normal image, and thus achieving luminance-invariant feature extraction. To preserve semantic meaning in the transformed features, a training policy has been introduced for object detection and feature transformation in various domains. The proposed method achieves state-of-the-art performance with a 9.95 improvement in average precision without incurring additional computational cost.

Loess landslides detection via a partially supervised learning and improved Mask-RCNN with multi-source remote sensing data

Loess landslides pose a severe threat of destruction, and detecting them is crucial for minimizing their impact on society. They typically consist of wind-deposited clay and silt, which makes them challenging to detect using conventional methods. Techniques like visual interpretation and field surveys are the most useful, yet these methods can be laborious, expensive, and require a certain level of prior knowledge. Furthermore, remote sensing approaches face the challenge of distinguishing between natural erosion and landslides. Recently, deep learning for landslide detection has the potential to speed up and improve detection accuracy. Nonetheless, deep learning models require large amounts of labeled data and the development of robust algorithms capable of extracting meaningful features from remote sensing data. In this article, a novel approach is introduced to improve the Mask Regional Convolutional Neural Network (Mask-RCNN) algorithm for accurately detecting landslides when the number of available segmentation mask samples is limited. Specifically, A novel loess landslides dataset is established using high-resolution remote sensing images in Gansu Province. In the context of partially supervised learning, a neural network branch containing a weight transfer function is developed to capture mask information from bounding boxes. Additionally, a mask-scoring network block is used to learn the quality of predicted instance masks. Our modified algorithm achieves an average precision improvement of 20.7% compared to the original Mask R-CNN algorithm in small landslide detection. The mask IoU threshold value of 0.5 is used to estimate the average accuracy higher than 0.75. The average precision of the segmentation mask is improved by 16.7% in test set. By proposing a solution that can achieve accurate landslide detection while using limited labeled samples, this study makes a valuable contribution to the application of deep learning in the domains of remote sensing and landslide detection.

SIRN: An iterative reasoning network for transmission lines based on scene prior knowledge

High-accuracy positioning and identification of transmission line components is the premise for their status detection and fault diagnosis. However, due to the limitations of imbalance object scale and distribution in aerial images, the problem of detecting dense-tiny objects still needs to be solved. Considering the prior knowledge of fixed connection scenes, we proposed a novel method named scene iterative reasoning network (SIRN) based on self-adaptive clustering and local scene knowledge. It consists of the coarse-grained detector (CGD), scene self-adaptive clustering (SSAC) subnetwork, and scene structure iterative reasoning (SSIR) subnetwork. The CGD was proposed to obtain global coarse detection results. Then, the SSAC subnetwork utilized unsupervised self-adaptive clustering to obtain accurate dense regions based on the CGD results. Finally, the SSIR subnetwork was designed to extract and fuse the scene semantic structure information of local regions for realizing accurate dense-tiny object detection. In summary, the SIRN model utilizes an iterative inference strategy to detect dense regions accurately. By leveraging scene structure information, the model effectively transforms the challenge of detecting dense objects into an advantage, resulting in precise detection. The SIRN model is compatible with single-stage and two-stage object detection models, achieving a notable mean average precision (mAP) improvement ranging from 4.4% to 12.2% compared to the baseline models. Compared to other state-of-the-art object detection models, SIRN demonstrates significant advantages in accuracy and error rate metrics. In addition, qualitative and quantitative experiments show that the SIRN model considerably enhanced the detection of dense-tiny objects. The code is available at https://github.com/CharmingWang/SIRN.

YOLOv7-CSAW for maritime target detection.

The issue of low detection rates and high false negative rates in maritime search and rescue operations has been a critical problem in current target detection algorithms. This is mainly due to the complex maritime environment and the small size of most targets. These challenges affect the algorithms' robustness and generalization. We proposed YOLOv7-CSAW, an improved maritime search and rescue target detection algorithm based on YOLOv7. We used the K-means++ algorithm for the optimal size determination of prior anchor boxes, ensuring an accurate match with actual objects. The C2f module was incorporated for a lightweight model capable of obtaining richer gradient flow information. The model's perception of small target features was increased with the non-parameter simple attention module (SimAM). We further upgraded the feature fusion network to an adaptive feature fusion network (ASFF) to address the lack of high-level semantic features in small targets. Lastly, we implemented the wise intersection over union (WIoU) loss function to tackle large positioning errors and missed detections. Our algorithm was extensively tested on a maritime search and rescue dataset with YOLOv7 as the baseline model. We observed a significant improvement in the detection performance compared to traditional deep learning algorithms, with a mean average precision (mAP) improvement of 10.73% over the baseline model. YOLOv7-CSAW significantly enhances the accuracy and robustness of small target detection in complex scenes. This algorithm effectively addresses the common issues experienced in maritime search and rescue operations, specifically improving the detection rates and reducing false negatives, proving to be a superior alternative to current target detection algorithms.

MSTA-SlowFast: A Student Behavior Detector for Classroom Environments.

Detecting students' classroom behaviors from instructional videos is important for instructional assessment, analyzing students' learning status, and improving teaching quality. To achieve effective detection of student classroom behavior based on videos, this paper proposes a classroom behavior detection model based on the improved SlowFast. First, a Multi-scale Spatial-Temporal Attention (MSTA) module is added to SlowFast to improve the ability of the model to extract multi-scale spatial and temporal information in the feature maps. Second, Efficient Temporal Attention (ETA) is introduced to make the model more focused on the salient features of the behavior in the temporal domain. Finally, a spatio-temporal-oriented student classroom behavior dataset is constructed. The experimental results show that, compared with SlowFast, our proposed MSTA-SlowFast has a better detection performance with mean average precision (mAP) improvement of 5.63% on the self-made classroom behavior detection dataset.

DMS-YOLOv5: A Decoupled Multi-Scale YOLOv5 Method for Small Object Detection

Small objects detection is a challenging task in computer vision due to the limited semantic information that can be extracted and the susceptibility to background interference. In this paper, we propose a decoupled multi-scale small object detection algorithm named DMS-YOLOv5. The algorithm incorporates a receptive field module into the feature extraction network for better focus on low-resolution small objects. The coordinate attention mechanism, which combines spatial and channel attention information, is introduced to reduce interference from background information and enhance the network’s attention to object information. A detection layer tailored to small-sized objects is added to compensate for the loss of small object information in multiple downsampling operations, greatly improving the detection capability of small objects. Next, The decoupled network is introduced into the detection head network for branch processing of classification and bounding box regression tasks. Finally, the bounding box loss function is improved to alleviate missed detection problems caused by the concentration of small objects and mutual occlusion between objects. The improved method achieved a mean average precision improvement of 12.1% on VisDrone2019-DET dataset compared to the original method. In comparison experiments with similar methods, our proposed method also demonstrated good performance, validating its effectiveness.

Learning for Vehicle-to-Vehicle Cooperative Perception Under Lossy Communication

Deep learning has been widely used in intelligent vehicle driving perception systems, such as 3D object detection. One promising technique is Cooperative Perception, which leverages Vehicle-to-Vehicle (V2V) communication to share deep learning-based features among vehicles. However, most cooperative perception algorithms assume ideal communication and do not consider the impact of Lossy Communication (LC), which is very common in the real world, on feature sharing. In this paper, we explore the effects of LC on Cooperative Perception and propose a novel approach to mitigate these effects. Our approach includes an LC-aware Repair Network (LCRN) and a V2V Attention Module (V2VAM) with intra-vehicle attention and uncertainty-aware inter-vehicle attention. We demonstrate the effectiveness of our approach on the public OPV2V dataset (a digital-twin simulated dataset) using point cloud-based 3D object detection. Our results show that our approach improves detection performance under lossy V2V communication. Specifically, our proposed method achieves a significant improvement in Average Precision compared to the state-of-the-art cooperative perception algorithms, which proves the capability of our approach to effectively mitigate the negative impact of LC and enhance the interaction between the ego vehicle and other vehicles. The code is available at: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/jinlong17/V2VLC</uri>

LightCSPNet: A Lightweight Network for Image Classification and Objection Detection

In recent years, computer vision and convolutional neural networks have been gradually applied in embedded devices. However, due to the limitation of hardware, the inference speed of many high-precision algorithms is very slow, which requires high performance hardware. In this study, a lightweight network called LightCSPNet is proposed for image classification and object detection. LightCSPNet is built by stacking four identical modules, each of which has adopted an improved CSP (Cross-Stage-Partial-connections) structure for channel number expansion. The special inverse residual structure is constructed for feature extraction, and the transformer modules are added in the proposed model. In this study, the typical defect detection in industry is adopted as testing platform, and a defect dataset consisting of 12 categories including cloth, road, bridge, steel and etc., was constructed for image classification. Compared with MobileNetV3, our model has almost the same accuracy, but the number of parameters and GFLOPs (Giga Floating-point Operations Per Second) have been, respectively, reduced to 88% and 36% for ImageNet100 and the dataset we built. In addition, compared with MobileNetV2 and MobileNetV3 for VOC2012 dataset in object detection, LightCSPNet obtained 0.4% and 0.6% mAP (Mean Average Precision) improvement respectively, and the inference speed on CPU was twice as fast.

Multitemporal greenhouse mapping for high-resolution remote sensing imagery based on an improved YOLOX

Plastic greenhouses (PGs), as safeguard agricultural facilities, have changed the cultivation structure of agriculture, to a certain extent, and can seriously impact the environment. Exploring the temporal characteristics of greenhouses in terms of quantity and area is significant for sustainable agricultural development and environmental protection. Pixel or object-based large-scale greenhouse mapping has difficulty meeting the demand for the simultaneous extraction of greenhouse quantity and area. Multibranch task model can provide the number and area information of greenhouse. However, there are various labeling methods and model between different branches, which lead to the deviation between quantity and area results. This paper proposes a model that is based on improved you only look once X (YOLOX) and kullback–leibler divergence (KLD), called YOLOX and KLD plus (YXLD + ), to obtain spatiotemporally consistent greenhouse mapping. The new model still has stable performance in dense scenes, with a 5.7 % improvement in average precision compared to BBAVectors. Finally, a multitemporal greenhouse map of Zhoukou City was obtained using imagery provided by Google Earth (GE). Overall, the distribution of greenhouses in Zhoukou City remained stable between 2012 and 2017. However, the number and area of greenhouses in local areas has increased rapidly in the past five years, with the number of greenhouses increasing by 47.08 % and the area expanding by 41.38 %. Greenhouses' distribution and changing characteristics concentrate on the gently sloping areas from the northwest to the southeast. These results provide an objective perspective on the development of the local agricultural cultivation structure, which is of significant importance for the promotion of agricultural modernization.

PCLUDA: A Pseudo-Label Consistency Learning- Based Unsupervised Domain Adaptation Method for Cross-Domain Optical Remote Sensing Image Retrieval

Recent advances in deep learning have dramatically improved the performance of content-based remote sensing image retrieval (CBRSIR) with the same distribution of training set (source domain) and test set (target domain). In fact, their distributions are inconsistent in most cases, which can lead to a dramatic decrease in retrieval performance. Currently, some unsupervised domain adaptation (DA) methods for other remote sensing applications have been proposed to eliminate the inconsistency. However, the current unsupervised DA methods do not make full use of the target domain’s distribution characteristics when delineating its decision boundary. This tends to degrade the cross-domain retrieval performance. In this article, a pseudo-label consistency learning-based unsupervised DA method (PCLUDA) is proposed for cross-domain CBRSIR. Our PCLUDA method minimizes the difference in probability distribution between the target domain and its perturbed output by a pseudo-label self-training and consistency regularization strategy, followed by adjusting the target domain’s decision boundaries to the low-density region. Besides, minimize class confusion (MCC) is introduced to reduce negative transfer caused by large intraclass variance of RSIs. Two cross-domain datasets with 12 cross-domain scenarios are constructed based on six open access datasets to measure DA methods. Experimental results show that our PCLUDA method achieves superior retrieval performances with average retrieval precision improvement by 4.9%–32.3% compared with eight state-of-the-art DA approaches in complex cross-domain scenarios. Furthermore, other experimental results indicate that our PCLUDA can also reach optimal retrieval performances in different kinds of deep learning networks [i.e., vision transformer (ViT) and convolutional neural networks (CNNs)].

MCL: A Contrastive Learning Method for Multimodal Data Fusion in Violence Detection

Multimodal learning among video and audio has shown significant performance improvement in violence detection. However, video and audio do not contribute consistently, and the video modality tends to dominate when determining whether a certain scene contains violent events. In fact, a few recent multimodal learning methods for violence detection do not fully consider data differences between various modalities, which lead to optimization imbalance problem during training, and ultimately result in insufficient performance. To address this issue, we propose a Multimodal Contrastive Learning (MCL) method to make full use of video and audio information for violence detection. In specific, to avoid the video modality dominating the model training, we design a multi-encoder framework to perform task-driven feature encoding on video and audio respectively. To reduce information loss during multimodal fusion, we introduce a contrastive learning task to capture semantically consistent representations. We conduct extensive experiments on XD-Violence dataset, showing that our proposed MCL achieves an average precision improvement of 2.34% against the state-of-the-art baseline.

Accurate Ship Detection Using Electro-Optical Image-Based Satellite on Enhanced Feature and Land Awareness

This paper proposes an algorithm that improves ship detection accuracy using preprocessing and post-processing. To achieve this, high-resolution electro-optical satellite images with a wide range of shape and texture information were considered. The developed algorithms display the problem of unreliable detection of ships owing to clouds, large waves, weather influences, and shadows from large terrains. False detections in land areas with image information similar to that of ships are observed frequently. Therefore, this study involves three algorithms: global feature enhancement pre-processing (GFEP), multiclass ship detector (MSD), and false detected ship exclusion by sea land segmentation image (FDSESI). First, GFEP enhances the image contrast of high-resolution electro-optical satellite images. Second, the MSD extracts many primary ship candidates. Third, falsely detected ships in the land region are excluded using the mask image that divides the sea and land. A series of experiments was performed using the proposed method on a database of 1984 images. The database includes five ship classes. Therefore, a method focused on improving the accuracy of various ships is proposed. The results show a mean average precision (mAP) improvement from 50.55% to 63.39% compared with other deep learning-based detection algorithms.

Improving the Accuracy of an R-CNN-Based Crack Identification System Using Different Preprocessing Algorithms

The accurate intelligent identification and detection of road cracks is a key issue in road maintenance, and it has become popular to perform this task through the field of computer vision. In this paper, we proposed a deep learning-based crack detection method that initially uses the idea of image sparse representation and compressed sensing to preprocess the datasets. Only the pixels that represent the crack features remain, while most pixels of non-crack features are relatively sparse, which can significantly improve the accuracy and efficiency of crack identification. The proposed method achieved good results based on the limited datasets of crack images. Various algorithms were tested, namely, linear smooth, median filtering, Gaussian smooth, and grayscale threshold, where the optimal parameters of the various algorithms were analyzed and trained with faster regions with convolutional neural network features (faster R-CNN). The results of the experiments showed that the proposed method has good robustness, with higher detection efficiency in the presence of, for example, road markings, shallow cracks, multiple cracks, and blurring. The result shows that the improvement of mean average precision (mAP) can reach 5% compared with the original method.

Improving Deforestation Detection on Tropical Rainforests Using Sentinel-1 Data and Convolutional Neural Networks

Detecting early deforestation is a fundamental process in reducing forest degradation and carbon emissions. With this procedure, it is possible to monitor and control illegal activities associated with deforestation. Most regular monitoring projects have been recently proposed, but most of them rely on optical imagery. In addition, these data are seriously restricted by cloud coverage, especially in tropical environments. In this regard, Synthetic Aperture Radar (SAR) is an attractive alternative that can fill this observational gap. This work evaluated and compared a conventional method based on time series and a Fully Convolutional Network (FCN) with bi-temporal SAR images. These approaches were assessed in two regions of the Brazilian Amazon to detect deforestation between 2019 and 2020. Different pre-processing techniques, including filtering and stabilization stages, were applied to the C-band Sentinel-1 images. Furthermore, this study proposes to provide the network with the distance map to past-deforestation as additional information to the pair of images being compared. In our experiments, this proposal brought up to 4% improvement in average precision. The experimental results further indicated a clear superiority of the DL approach over a time series-based deforestation detection method used as a baseline in all experiments. Finally, the study proved the benefits of pre-processing techniques when using detection methods based on time series. On the contrary, the analysis revealed that the neural network could eliminate noise from the input images, making filtering innocuous and, therefore, unnecessary. On the other hand, the stabilization of the input images brought non-negligible accuracy gains to the DL approach.