Mask R-CNN Algorithm Research Articles

Mask R-CNN and Centroid Tracking Algorithm to Process UAV Based Thermal-RGB Video for Drylot Cattle Heat Stress Monitoring

This study developed and evaluated an algorithm for processing thermal-RGB video feeds captured by an unmanned aerial vehicle (UAV) to automate heat stress monitoring in cattle housed in the drylots. The body surface temperature (BST) of individual cows was used as an indicator of heat stress. UAV data were collected using RGB and thermal infrared imagers, respectively, at 2 and 6.67 cm per pixel spatial resolution in Spring 2023 (dataset-1) and Summer 2024 (dataset-2). Study sites were two commercial drylots in Washington State. The custom algorithms were developed to: (1) detect and localize individual cows using a Mask R-CNN-based instance segmentation model combined with centroid tracking; and (2) extract BST by averaging the thermal-imagery pixels for each of the segmented cows. The algorithm showed higher detection accuracy with RGB images as input (F1 score: 0.89) compared to thermal (F1 score: 0.64). BST extraction with combined RGB and thermal imaging approach required corrections for alignment problems associated with differences in optics, imaging field of view, resolution, and lens properties. Consequently, thermal imaging-only approach was adopted for assessing real-time cow localization and BST estimation. Operating at one frame per second, algorithm successfully detected 72.4% and 81.65% of total cows in video frames from dataset-1 (38 s) and -2 (48 s), respectively. The mean absolute difference between algorithm output and ground truth (BSTGT) was 2.1 °C (dataset-1) and 3.3 °C (dataset-2), demonstrating satisfactory performance. With further refinements, this approach could be a viable tool for real-time heat stress monitoring in large-scale drylot production systems.

Pipeline multi-type high consequence area identification based on mask R-CNN with fused attention mechanisms

Identification of high consequence areas is an important task in pipeline integrity management. However, traditional identification methods are generally characterized by low efficiency, high cost and low accuracy. For this reason, this paper proposes a recognition method based on the improved algorithm Mask Region-based Convolutional Neural Network. Coordinate attention mechanism module is introduced into the traditional Mask R-CNN algorithm to improve the recognition accuracy and reduce the training time. For the identification results, GIS tools are utilized to establish high consequence zones along both sides of the pipeline, and the grade and scope of the high consequence zones are determined according to relevant specifications.In this paper, this method is used to identify the high-consequence area of a pipeline section in Guangdong Province, the results show that: 1, the improved algorithm in the identification of densely populated, geologic hazards, flammable and explosive high consequence zones of the average accuracy of the identification of 1.7%, 3.4%, 3.9%. 2, The method in this paper identifies 8 more building elements and 0.311 more kilometers of pipeline mileage compared to traditional identification methods. The method of this paper can provide a reference for the early identification of high consequence areas of pipelines.

Detection of Straw Coverage under Conservation Tillage Based on an Improved Mask Regional Convolutional Neural Network (Mask R-CNN)

Conservation tillage, a crucial method for protecting soil fertility, depends heavily on maintaining adequate straw coverage. The current method of straw coverage detection relies primarily on manual measurement, which is both time-consuming and laborious. This paper introduces a novel straw coverage detection approach based on an improved mask regional convolutional neural network (Mask R-CNN) algorithm. Several images of wheat straw-covered fields were taken, and the dataset was augmented using techniques like image inversion, contrast enhancement, Gaussian noise addition, and translation after cropping the original images. These fields use a crop rotation cycle of wheat and corn. Subsequently, the straw images were annotated using the Labelme annotation tool to obtain the available straw instance segmentation dataset. The Mask R-CNN algorithm was improved by refining the mask generation network structure through a multi-feature fusion strategy, which interweaves features from both the encoder and the mask generation network, enhancing the model’s ability to capture detailed and shape information of the straw. Lastly, using the mask information output by the improved Mask R-CNN algorithm, the straw coverage was calculated by counting the proportion of pixels within each segmented region. In the results, compared to the original Mask R-CNN algorithm, our improved Mask R-CNN algorithm achieved an average improvement of 7.8% in segmentation accuracy, indicating that the improved Mask R-CNN algorithm offers superior segmentation performance. Thus, the new algorithm can achieve straw coverage detection with higher accuracy and can provide a reference for other agricultural applications.

Application of Mask R-CNN Algorithm for Apple Detection and Semantic Segmentation

Application of Mask R-CNN Algorithm for Apple Detection and Semantic Segmentation

Study on Fault Monitoring Technology of Photovoltaic Panel Based on Thermal Infrared and Optical Remote Sensing

Abstract. Rapid access to the operating status of Photovoltaic (PV) panels and troubleshooting can save management and maintenance costs for the development of PV power plants, which is important for PV power plant management and power generation capacity assurance. The use of remote sensing technology to identify the faults of photovoltaic panels has developed rapidly, however, current research usually relies only on a single optical data source to identify and count the area of PV panels in a PV electric field, although there are literature on PV panel fault detection, only the surface fault identification of PV panels is tested, while the internal faults (such as panel bad points or bad lines) cannot be identified because of the limitations of optical remote sensing. In this paper, a photovoltaic panel fault monitoring technology based on multi-source remote sensing is proposed. The optical and thermal infrared hybrid data combined with deep learning technology are used to achieve rapid and accurate fault identification and localization of PV panel arrays. It can not only automatically identify PV panels that are obscured by dust and foreign objects, but also locate PV panels that have bad dots or bad lines, which greatly improves the ability and effectiveness of remote sensing PV panel fault monitoring. The high-resolution unmanned air vehicle (UAV) optical image and thermal infrared image are applied in this experiment. The Mask RCNN algorithm is used to accurately locate and number the photovoltaic panel of the optical image. Then, the fault scene classification model is established for the multi-type fault characteristics of the optical image and thermal infrared image within the panel range, so as to identify five types of faults, such as dust cover, branch cover, bird droppings cover, internal bad points and bad lines of PV panel, which effectively solves the problem that the single optical remote sensing image cannot identify the internal component faults of the photovoltaic panel under normal conditions.

A methodology for phase characterization in pellet feed using digital microscopy and deep learning

This work proposes a new method for characterizing pellet feed, employing Deep Learning (DL) and Convolutional Neural Networks (CNNs). The main minerals in the composition of the studied pellet feed are hematite, magnetite, goethite, and quartz. Over time, several characterization methodologies have been developed that use Digital Microscopy and Image Analysis tools. The greatest difficulties in this characterization lie in differentiating the textures of hematite particles, the different shapes of their crystals, or discriminating between quartz and resin in reflected light optical microscopy images. This work proposes a mineral characterization methodology based on the Mask R-CNN algorithm. The goal is to perform instance segmentation, that is, to identify, classify, and segment objects in the images. Two DL models were combined: the BF Model performs instance segmentation for the compact, porous, martite, and goethite classes in images obtained in Bright Field mode, and the CPOL Model uses images acquired in Circularly Polarized Light to segment the monocrystalline, polycrystalline, and martite classes. An F1-score of around 80% was obtained for the BF Model and around 90% for the CPOL Model. The results were promising and can be improved as the training image database increases.

A deep semantic network-based image segmentation of soybean rust pathogens.

Asian soybean rust is a highly aggressive leaf-based disease triggered by the obligate biotrophic fungus Phakopsora pachyrhizi which can cause up to 80% yield loss in soybean. The precise image segmentation of fungus can characterize fungal phenotype transitions during growth and help to discover new medicines and agricultural biocides using large-scale phenotypic screens. The improved Mask R-CNN method is proposed to accomplish the segmentation of densely distributed, overlapping and intersecting microimages. First, Res2net is utilized to layer the residual connections in a single residual block to replace the backbone of the original Mask R-CNN, which is then combined with FPG to enhance the feature extraction capability of the network model. Secondly, the loss function is optimized and the CIoU loss function is adopted as the loss function for boundary box regression prediction, which accelerates the convergence speed of the model and meets the accurate classification of high-density spore images. The experimental results show that the mAP for detection and segmentation, accuracy of the improved algorithm is improved by 6.4%, 12.3% and 2.2% respectively over the original Mask R-CNN algorithm. This method is more suitable for the segmentation of fungi images and provide an effective tool for large-scale phenotypic screens of plant fungal pathogens.

Body weight prediction in crossbred pigs from digital images using computer vision

The development of technologies made it possible to capture digital images by sensors in order to obtain morphometric measurements that can be used in the development of mathematical models for the prediction of body weight (BW) in animals. The objective was to develop mathematical models to predict body weight in crossbred pigs from morphometric measurements obtained with two-dimensional digital images. Data collection was carried out in the swine sector of IF Goiano, Urutai campus, in 52 crossbred pigs housed in the finishing phase. The animals were randomly weighed and filmed by the Microsoft Kinect® v1 sensor camera in the dorsal view. Subsequently, they were conducted for the measurement of manual morphometric measurements, without restraint, being dorsal length, chest width and flank width. The morphometric measurements predicted from the dorsal images were dorsal length, chest width, flank width, dorsal area and perimeter. Through Mask R-CNN algorithm trains the network by input images, was compute the area of the evaluated object in pixels, and thus, can correlate with measures of interest. The Mask R-CNN algorithm were performed with the Collaboratory Google using Phyton 3.0. Mathematical models were developed using multiple linear regression (MLR) and support vector regression (SVR) methodologies for body weight prediction. All Pearson's correlations between real body weight and real and predicted morphometric measurements by digital images were moderate to high magnitude positive and statistically significant (P<0.05). The difference between the real and predicted BW by the SVR was -1.74 kg (R² = 91%) and -2.39 kg (R² = 88%), respectively. The MLR model with real morphometrics measurements explained 60% of BW variance and estimated a BW 18.76 kg below from the real BW mean. The MLR model with predicted measurements explained 53% of BW variance and estimated a BW 22.51 kg below from the real BW mean. The mathematical model developed from SVR has the potential to estimate the body weight of crossbred pigs using morphometric measurements predicted by digital images such as dorsal length, width, dorsal area and perimeter.

Fish age reading using deep learning methods for object-detection and segmentation

Abstract Determination of individual age is one essential step in the accurate assessment of fish stocks. In non-tropical environments, the manual count of ring-like growth patterns in fish otoliths (ear stones) is the standard method. It relies on visual means and individual judgment and thus is subject to bias and interpretation errors. The use of automated pattern recognition based on machine learning may help to overcome this problem. Here, we employ two deep learning methods based on Convolutional Neural Networks (CNNs). The first approach utilizes the Mask R-CNN algorithm to perform object detection on the major otolith reading axes. The second approach employs the U-Net architecture to perform semantic segmentation on the otolith image in order to segregate the regions of interest. For both methods, we applied a simple postprocessing to count the rings on the output masks returned, which corresponds to the age prediction. Multiple benchmark tests indicate the promising performance of our implemented approaches, comparable to recently published methods based on classical image processing and traditional CNN implementation. Furthermore, our algorithms showed higher robustness compared to the existing methods, while also having the capacity to extrapolate missing age groups and to adapt to a new domain or data source.

Using mask R-CNN to rapidly detect the gold foil shedding of stone cultural heritage in images

As immovable stone cultural heritage is kept in the open air, they are more susceptible to damage, and damage detection is very important for the protection and restoration of cultural heritage. This is especially true for gold-overlaid stone cultural heritage, which is usually more complicated than ordinary stone carvings. However, the detection of cultural heritage damages is mainly based on expert visual inspection, which is often subjective, time-consuming, and laborious. This paper uses the Mask R-CNN algorithm to rapidly and accurately detect the gold foil shedding of stone cultural heritage through two-dimensional images. The research data are from the high-precision images of the Dazu Thousand-Hand Bodhisattva Statue (World Heritage, UNESCO) in Chongqing, China. After cleaning and augmentation, 1900 images are input into Mask R-CNN model for training. Finally, the average precision value (AP) for detecting gold foil shedding is found to be 0.967. In order to test the performance of the model, the new images that do not participate in the training period are used, and it is found that the model can still accurately detect the gold foil shedding even if there are interference factors. This is the first attempt to detect the damages of gold-overlaid stone cultural heritage based on a deep learning algorithm, and it has achieved good results.

Acoustic neuroma classification algorithm based on mask region convolution neural network

ObjectiveAcoustic neuroma has similarities with other intracranial tumors in imaging manifestations and location of incidence, and misdiagnosis often occurs in clinical practice. This paper uses a mask region convolution neural network (Mask RCNN) to classify acoustic neuromas. MethodsThe T1WI-SE sequence MRI images of 120 patients with acoustic neuroma in our hospital were collected. Based on preprocessing, the improved feature pyramid networks (FPN) algorithm and Mask RCNN comprehensive training were conducted, and the classification effects of different networks were compared. ResultsThe accuracy of the Mask RCNN classification model of ResNet101 network was 0.92, the recall rate was 0.86, the specificity was 0.89, and the mean average precision (mAP) was 0.91. ConclusionThe classification model based on Mask RCNN algorithm has a good effect on the differentiation and classification of acoustic neuroma.

Deep learning-based image analysis for in situ microscopic imaging of cell culture process

Mammalian cell culture is an important bioprocess that directly affects the quality and yield of biopharmaceuticals. Traditionally, condition monitoring of the operation is based on sampling periodically and off-line analysis, which is labor intensive, time consuming, and causing time delays. In this work, in situ microscope is investigated for on-line real-time monitoring of the culture process of Chinese hamster ovary cells with focus on investigation of deep learning-based Mask R-CNN algorithm for image analysis. The model is trained by 184 images with 183,040 cells using data augmentation methods and transfer learning technique. Mask R-CNN segmented the clustered cells more effectively than the conventional one combining edge detection, intensity thresholding, and advanced watershed method as well as the multi-scale edge detection method. Its Dice score, accuracy, precision, sensitivity, F1 score, specificity, and relative volume difference reach 0.862, 0.945, 0.901, 0.827, 0.862, 0.977, and 0.082, respectively. The evolution of geometrical features of cells were further analyzed, including equivalent diameter, circularity, aspect ratio, and eccentricity. The result demonstrated the great potential of deep learning technology in analysis of on-line images for optimization and control of the cell culture process.

A Novel Road Crack Detection Technology Based on Deep Dictionary Learning and Encoding Networks

Road crack detection is an important indicator of road detection. In real life, it is very meaningful work to detect road cracks. With the rapid development of science and technology, especially computer science and technology, quite a lot of methods have been applied to crack detection. Traditional detection methods rely on manual identification, which is inefficient and prone to errors. In addition, the commonly used image processing methods are affected by many factors, such as illumination, road stains, etc., so the results are unstable. In the research on pavement crack detection, many research studies mainly focus on the recognition and classification of cracks, lacking the analysis of the specific characteristics of cracks, and the feature values of cracks cannot be measured. Starting from the deep learning method in computer science and technology, this paper proposes a road crack detection technology based on deep learning. It relies on a new deep dictionary learning and encoding network DDLCN, establishes a new activation function MeLU, and adopts a new differentiable calculation method. The technology relies on the traditional Mask-RCNN algorithm and is implemented after improvement. In the comparison of evaluation indicators, the values of recall, precision, and F1-score reflect certain superiority. Experiments show that the proposed method has good implementability and performance in road crack detection and crack feature measurement.

110 Evaluation of Computer Vision to Analyze Beef Cattle Feeding Behavior

Abstract Leveraging computer vision (CV) methods to study pen-fed cattle feeding behavior presents several advantages, including monitoring animal health and identifying feed efficient animals. This research employed a region-based convolutional neural network (RCNN) combined with the common objects in context (COCO) dataset for automatic livestock recognition using CV techniques in experimental feedlot pens. Thirty Angus-influenced steers were allocated in one pen with four automated feed intake systems (AFIS; Vytelle SENSE). CV data were recorded during daylight hours using a webcam (Microsoft LifeCam Cinema) with a resolution of 1280x720 pixels at ten frames per second connected to a video surveillance camera software (Contaware, Switzerland). The CV dataset obtained by the cameras, Figure 1, was benchmarked with feeding behavior data obtained from observed annotations (OA) and AFIS. The CV model utilized was a Mask-RCNN with pre-trained weights on the COCO dataset. The Mask RCNN algorithm can identify and locate multiple objects, objects of different scales, and overlapping objects within an image. The COCO dataset is a large and diverse set of annotated images that can be used for training and evaluating object detection models. The fully trained Mask-RCNN model takes each video frame as input and outputs the pixel segmentation of each object detected. For the CV analyses, each animal bounding box was associated with a score threshold, and each pixel within was assigned a probability threshold; both ranging from 0 to 100. The higher the score and probability thresholds are, the fewer the boxes and pixels will be marked as the animal boxing shape, resulting in less feeding detection from the CV system. In contrast, a low score and probability thresholds produce many predictions, including false detections, which might be misleading. Therefore, a dataset analysis was conducted to optimize the best threshold combination. We evaluated the accuracy and precision of the CV model compared with OA and AFIS by analyzing a dataset section of feeding events of one feed bunk using a 0.5 score and 0.5 probability thresholds. We obtained the values for the true positive (TP), false positive (FP), true negative (TN), and false negative (FN) to calculate the accuracy = [TP+TN]/[TP+TN+FP+FN], and precision = [TP]/[TP+FP], Table 1. As a result, the CV model achieved 95.39% and 99.82% accuracy and 93.60% and 99.90% precision when compared with OA and AFIS, respectively. Our findings showcase the promising capabilities of employing the Mask-RCNN algorithm and COCO datasets for detecting beef cattle feeding behavior in feedlot pens. This approach holds significant value for the cattle industry by enabling precise monitoring and analysis of individual animal behavior within feedlot pens, facilitating early detection of behavioral abnormalities.

Mask R-CNN–Based Landslide Hazard Identification for 22.6 Extreme Rainfall Induced Landslides in the Beijiang River Basin, China

Landslides triggered by extreme precipitation events pose a significant threat to human life and property in mountainous regions. Therefore, accurate identification of landslide locations is crucial for effective prevention and mitigation strategies. During the prolonged heavy rainfall events in Guangdong Province between 21 May and 21 June 2022, shallow and clustered landslides occurred in the mountainous regions of the Beijiang River Basin. This research used high-resolution satellite imagery and integrated the Mask R-CNN algorithm model with spectral, textural, morphological and physical characteristics of landslides in remote sensing imagery, in addition to landslide-influencing factors and other constraints, to interpret the landslides induced by the event through remote sensing techniques. The detection results show that the proposed methodology achieved a high level of accuracy in landslide identification, with a precision rate of 81.91%, a recall rate of 84.07% and an overall accuracy of 87.28%. A total of 3782 shallow landslides were detected, showing a distinct clustered distribution pattern. The performance of Mask R-CNN, Faster-CNN, U-Net and YOLOv3 models in landslide identification was further compared, and the effects of setting the rotation angle and constraints on the identification results of the Mask R-CNN algorithm model were investigated. The results show that each model improves the evaluation indices, but the Mask R-CNN model has the best detection performance; the rotation angle can effectively improve the generalization ability and robustness of the model, and the landslide-inducing factor data and texture feature sample data are the best for landslide identification. The research results provide valuable references and technical support for deepening our understanding of the distribution patterns of rainfall-triggered shallow and cluster landslides in the Beijiang River Basin.

New technologies for UAV navigation with real-time pattern recognition

The study is dedicated to the investigation and analysis of real-time navigation technologies for UAVs utilizing image scanning and object recognition. This research employs the DJI Inspire 2 drone alongside MapBotix software, integrating adapted algorithms such as Faster R-CNN and MaskRCNN. During the flight of the DJI Inspire 2 drone over the Banska Bystrica region, information concerning object recognition and change detection within the studied area was collected. A 40-second scanning process documented two alterations, one of which occurred at the 10th second. Subsequent stages of image recognition were executed utilizing scaling techniques, thus ensuring the accuracy of data extracted from captured images. This research, in particular, pioneers the exploration of real-time UAV navigation technologies integrated with image recognition. The application of a customized MaskRCNN algorithm, comprising 9 convolutional layers, 4 max-pooling layers, and 1 detection layer, for analyzing scanned object images signifies a novel approach in this domain.

A Deep Learning Network for Individual Tree Segmentation in UAV Images with a Coupled CSPNet and Attention Mechanism

Accurate individual tree detection by unmanned aerial vehicles (UAVs) is a critical technique for smart forest management and serves as the foundation for evaluating ecological functions. Existing object detection and segmentation methods, on the other hand, have reduced accuracy when detecting and segmenting individual trees in complicated urban forest landscapes, as well as poor mask segmentation quality. This study proposes a novel Mask-CSP-attention-coupled network (MCAN) based on the Mask R-CNN algorithm. MCAN uses the Cross Stage Partial Net (CSPNet) framework with the Sigmoid Linear Unit (SiLU) activation function in the backbone network to form a new Cross Stage Partial Residual Net (CSPResNet) and employs a convolutional block attention module (CBAM) mechanism to the feature pyramid network (FPN) for feature fusion and multiscale segmentation to further improve the feature extraction ability of the model, enhance its detail information detection ability, and improve its individual tree detection accuracy. In this study, aerial photography of the study area was conducted by UAVs, and the acquired images were used to produce a dataset for training and validation. The method was compared with the Mask Region-based Convolutional Neural Network (Mask R-CNN), Faster Region-based Convolutional Neural Network (Faster R-CNN), and You Only Look Once v5 (YOLOv5) on the test set. In addition, four scenes—namely, a dense forest distribution, building forest intersection, street trees, and active plaza vegetation—were set up, and the improved segmentation network was used to perform individual tree segmentation on these scenes to test the large-scale segmentation ability of the model. MCAN’s average precision (AP) value for individual tree identification is 92.40%, which is 3.7%, 3.84%, and 12.53% better than that of Mask R-CNN, Faster R-CNN, and YOLOv5, respectively. In comparison to Mask R-CNN, the segmentation AP value is 97.70%, an increase of 8.9%. The segmentation network’s precision for the four scenes in multi-scene segmentation ranges from 95.55% to 92.33%, showing that the proposed network performs high-precision segmentation in many contexts.

Improved YOLOv8-Seg Network for Instance Segmentation of Healthy and Diseased Tomato Plants in the Growth Stage

The spread of infections and rot are crucial factors in the decrease in tomato production. Accurately segmenting the affected tomatoes in real-time can prevent the spread of illnesses. However, environmental factors and surface features can affect tomato segmentation accuracy. This study suggests an improved YOLOv8s-Seg network to perform real-time and effective segmentation of tomato fruit, surface color, and surface features. The feature fusion capability of the algorithm was improved by replacing the C2f module with the RepBlock module (stacked by RepConv), adding SimConv convolution (using the ReLU function instead of the SiLU function as the activation function) before two upsampling in the feature fusion network, and replacing the remaining conventional convolution with SimConv. The F1 score was 88.7%, which was 1.0%, 2.8%, 0.8%, and 1.1% higher than that of the YOLOv8s-Seg algorithm, YOLOv5s-Seg algorithm, YOLOv7-Seg algorithm, and Mask RCNN algorithm, respectively. Meanwhile, the segment mean average precision (segment mAP@0.5) was 92.2%, which was 2.4%, 3.2%, 1.8%, and 0.7% higher than that of the YOLOv8s-Seg algorithm, YOLOv5s-Seg algorithm, YOLOv7-Seg algorithm, and Mask RCNN algorithm. The algorithm can perform real-time instance segmentation of tomatoes with an inference time of 3.5 ms. This approach provides technical support for tomato health monitoring and intelligent harvesting.

Instance segmentation of upper aerodigestive tract cancer: site-specific outcomes.

To achieve instance segmentation of upper aerodigestive tract (UADT) neoplasms using a deep learning (DL) algorithm, and to identify differences in its diagnostic performance in three different sites: larynx/hypopharynx, oral cavity and oropharynx. A total of 1034 endoscopic images from 323 patients were examined under narrow band imaging (NBI). The Mask R-CNN algorithm was used for the analysis. The dataset split was: 935 training, 48 validation and 51 testing images. Dice Similarity Coefficient (Dsc) was the main outcome measure. Instance segmentation was effective in 76.5% of images. The mean Dsc was 0.90 ± 0.05. The algorithm correctly predicted 77.8%, 86.7% and 55.5% of lesions in the larynx/hypopharynx, oral cavity, and oropharynx, respectively. The mean Dsc was 0.90 ± 0.05 for the larynx/hypopharynx, 0.60 ± 0.26 for the oral cavity, and 0.81 ± 0.30 for the oropharynx. The analysis showed inferior diagnostic results in the oral cavity compared with the larynx/hypopharynx (p < 0.001). The study confirms the feasibility of instance segmentation of UADT using DL algorithms and shows inferior diagnostic results in the oral cavity compared with other anatomic areas.

A boundary optimization model of instance segmentation combined with wavelet transform on Buildings

Data driven deep learning methods have become the mainstream method of building extraction from remote sensing images. In this paper, deep learning algorithm is used to classify and extract buildings from remote sensing images of rural areas around the Great Wall in the suburbs of Beijing captured by unmanned aerial vehicles. Aiming at the shortcomings of the current mainstream instance segmentation algorithm Mask R-CNN in feature fusion and poor prediction of instance mask boundaries, this paper proposes a boundary optimization algorithm for building instance segmentation based on discrete wavelet transform. Firstly, the discrete wavelet transform is introduced into the segmentation task branch of Mask R-CNN algorithm to extract the low-frequency and high-frequency information of the real mask, in which the high-frequency information includes the boundary information. Secondly, the pixel by pixel prediction of the mask turns into the learning of the low-frequency and high-frequency information of the real mask. The learning of the high-frequency information helps the segmentation network to learn the boundary features better. Finally, using the reversibility of discrete wavelet transform, the low-frequency and high-frequency information of the learned mask is inversely transformed to reconstruct the final mask. The improved algorithm is evaluated on the dataset COCO, and applied to the automatic extraction of buildings. The DWT Mask R-CNN algorithm model achieved 70.2% segmentation accuracy and 71.4% detection accuracy, which were improved by 1% and 0.7% respectively compared with the Mask R-CNN and Cascade Mask R-CNN models. The experimental results show that the instance segmentation edge optimization algorithm combined with wavelet transform has achieved better results on the segmentation boundary, improved the poor effect of mask edge detection and achieved higher detection accuracy, and can accurately extract village buildings.