Anchor Frames Research Articles

Research on Transformer Equipment Identification Based on Improved Fast RCNN Algorithm

Because of the problems such as the long work cycle and manual efficiency faced by the traditional manual inspection and video monitoring inspection methods of substations, this paper proposes an electrical equipment identification algorithm that improves the Fast RCNN network to realize automatic identification and positioning of electrical equipment. We used affine transformation, background fusion, hue and saturation change, clipping and filling methods to enrich images and change the feature extraction network from VGG16 to ResNet50 with residual structure so that the feature extraction network can focus on specific anchor frames, and then used K-means clustering algorithm in the regional recommendation network to increase the convergence speed of the model, in this way, the improved anchoring height-to-width ratio is more suitable for detecting the image of power equipment. The experimental simulation shows that the average detection accuracy of the improved model is better than that of the original model, with the average accuracy mAP reaching 90.04%, and the image detection speed reaching 11 frames/second.

VV-YOLO: A Vehicle View Object Detection Model Based on Improved YOLOv4

Vehicle view object detection technology is the key to the environment perception modules of autonomous vehicles, which is crucial for driving safety. In view of the characteristics of complex scenes, such as dim light, occlusion, and long distance, an improved YOLOv4-based vehicle view object detection model, VV-YOLO, is proposed in this paper. The VV-YOLO model adopts the implementation mode based on anchor frames. In the anchor frame clustering, the improved K-means++ algorithm is used to reduce the possibility of instability in anchor frame clustering results caused by the random selection of a cluster center, so that the model can obtain a reasonable original anchor frame. Firstly, the CA-PAN network was designed by adding a coordinate attention mechanism, which was used in the neck network of the VV-YOLO model; the multidimensional modeling of image feature channel relationships was realized; and the extraction effect of complex image features was improved. Secondly, in order to ensure the sufficiency of model training, the loss function of the VV-YOLO model was reconstructed based on the focus function, which alleviated the problem of training imbalance caused by the unbalanced distribution of training data. Finally, the KITTI dataset was selected as the test set to conduct the index quantification experiment. The results showed that the precision and average precision of the VV-YOLO model were 90.68% and 80.01%, respectively, which were 6.88% and 3.44% higher than those of the YOLOv4 model, and the model's calculation time on the same hardware platform did not increase significantly. In addition to testing on the KITTI dataset, we also selected the BDD100K dataset and typical complex traffic scene data collected in the field to conduct a visual comparison test of the results, and then the validity and robustness of the VV-YOLO model were verified.

Detecting diseases in apple tree leaves using FPN–ISResNet–Faster RCNN

ABSTRACT Apple leaf diseases typified by small disease spots are generally difficult to detect in images. This study proposes a deep learning model called the feature pyramid networks (FPNs) –inception squeeze-and-excitation ResNet (ISResNet)–Faster RCNN (region with convolutional neural network) model to improve the accuracy of detecting apple leaf diseases. Apple leaf diseases were identified, evaluated, and validated by using the FPN–ISResNet–Faster RCNN. The results were compared with those obtained by the single-shot multibox detector (SSD), Faster RCNN, and ISResNet–Faster RCNN. The detection accuracies obtained by using different feature extraction networks, positions and numbers of SE, inception modules, scales of FPN structures, and scales of anchor frames were also compared. The results showed that the values of average precision (AP), and APs with the thresholds of the intersection over union of 0.5 and 0.75 (AP50 and AP75), obtained from the FPN–ISResNet–Faster RCNN were 62.71%, 93.68%, and 70.94%, respectively, which are higher than those of the SSD, VGG–Faster RCNN, GoogleNet–Faster RCNN, ResNet50–Faster RCNN, ResNeXt–Faster RCNN, and ISResNet–Faster RCNN. FPN–ISResNet–Faster RCNN was shown to be able to detect diseases in apple leaves with high accuracy and generalizability.

An improved UAV target detection algorithm based on ASFF-YOLOv5s.

Object detection in drone-captured scenarios is a recent popular task. Due to the high flight altitude of unmanned aerial vehicle (UAV), the large variation of target scales, and the existence of dense occlusion of targets, in addition to the high requirements for real-time detection. To solve the above problems, we propose a real-time UAV small target detection algorithm based on improved ASFF-YOLOv5s. Based on the original YOLOv5s algorithm, the new shallow feature map is passed into the feature fusion network through multi-scale feature fusion to improve the extraction capability for small target features, and the Adaptively Spatial Feature Fusion (ASFF) is improved to improve the multi-scale information fusion capability. To obtain anchor frames for the VisDrone2021 dataset, we improve the K-means algorithm to obtain four different scales of anchor frames on each prediction layer. The Convolutional Block Attention Module (CBAM) is added in front of the backbone network and each prediction network layer to improve the capture capability of important features and suppress redundant features. Finally, to address the shortcomings of the original GIoU loss function, the SIoU loss function is used to accelerate the convergence of the model and improve accuracy. Extensive experiments conducted on the dataset VisDrone2021 show that the proposed model can detect a wide range of small targets in various challenging environments. At a detection rate of 70.4 FPS, the proposed model obtained a precision value of 32.55%, F1-score of 39.62%, and a mAP value of 38.03%, which improved 2.77, 3.98, and 5.1%, respectively, compared with the original algorithm, for the detection performance of small targets and to meet the task of real-time detection of UAV aerial images. The current work provides an effective method for real-time detection of small targets in UAV aerial photography in complex scenes, and can be extended to detect pedestrians, cars, etc. in urban security surveillance.

Lightweight and anchor-free frame detection strategy based on improved CenterNet for multiscale ships in SAR images

Ship detection using synthetic aperture radar (SAR) images has important applications in military and civilian fields, but the different sizes of the ship downgrade the detection accuracy of multiscale ships. Aiming at the problem of the poor accuracy and low efficiency of multiscale ship detection in complex scenes, this paper proposes a lightweight and anchor-free frame detection strategy for multiscale ships in SAR images. First, to deal with the problems of limited training samples, different sizes, attitudes, and angles of the ships in SAR images, a data augmentation strategy suitable for SAR images is adopted to expand the training space, followed by multiscale training to enhance the model generalization ability for multiscale ship detection. Second, a lightweight and anchor-free ship detection model based on the improved CenterNet is proposed, which abandons the dense anchor frame generation and extracts the key point of the ships for detection and positioning. Compared with the anchor frame-based detection method, this proposed detection model does not need to use the post-processing method to remove redundant anchor frames, and can accurately locate the center point of the ships with a better detection performance. Third, to reduce the model size and simplify the model parameters, a more lightweight network design is adopted in combination with the characteristics of SAR images. Hence, a residual network (ResNet) with fewer convolutional layers is constructed as the backbone network, and the cross-stage partial network (CSPNet) and spatial pyramid pooling (SPP) network are designed as the bottleneck network. The shallow ResNet can fully extract the SAR image features and reduce the training overfitting, and CSPNet and SPP can effectively combine the low-level image features to obtain the high-level features, reducing the model computation while at the same time enhancing the feature extraction ability. Finally, the evaluation index of the common objects in the context dataset is introduced, which can provide higher-quality evaluation results for ship detection accuracy and provide comprehensive evaluation indicators for multiscale ship detection. Experimental results show that the proposed strategy has the advantages of high detection efficiency, strong detection ability, and good generalization performance, which can achieve real-time and high-precision detection of the multiscale ship in complex SAR images.

Target Detection Network for Underwater Image Based on Adaptive Anchor Frame and Re-parameterization

For low target detection rate and inaccurate localization of most target detection networks in special underwater environments, we propose an underwater target detection algorithm with adaptive anchor frames. The adaptive anchor frame selection strategy differs from the traditional methods of manually designing anchor frames and obtaining anchor frames using the K-means clustering algorithm. Our method is based on the K-means clustering algorithm and consists of a grouping strategy and a dynamic K-value strategy. It can also automatically calculate the appropriate numbers of anchor boxes and layers of network output based on the characteristics of label data. And performance of the network is guaranteed while avoiding the computation caused by redundant structures. The simple but powerful RepVGG classification network is used as the backbone network. Therefore, the feature extraction capability of the network is enhanced and the detection performance of the network on fuzzy targets is improved. The experimental results show that the algorithm in this paper achieves better detection results in the underwater target detection task compared to other algorithms.

Pedestrian detection based on YOLOv3 multimodal data fusion

Multi-sensor fusion has essential applications in the field of target detection. Considering the current actual demand for miniaturization of on-board computers for driverless vehicles, this paper uses the multimodal data YOLOv3 (MDY) algorithm for pedestrian detection on embedded devices. The MDY algorithm uses YOLOv3 as the basic framework to improve pedestrian detection accuracy by optimizing anchor frames and adding small target detection branches. Then the algorithm is accelerated by using TensorRT technology to improve the real-time performance in embedded devices. Finally, a hybrid fusion framework is used to fuse the LIDAR point cloud data with the improved YOLOv3 algorithm to compensate for the shortcomings of a single sensor and improve the detection accuracy while ensuring speed. The improved YOLOv3 improves AP by 6.4% and speed by 11.3 FPS over the original algorithm. The MDY algorithm achieves better performance on the KITTI dataset. To further verify the feasibility of the MDY algorithm, an actual test was conducted on an unmanned vehicle with Jetson TX2 embedded device as the on-board computer within the campus scenario, and the results showed that the MDY algorithm achieves 90.8% accuracy under real-time operation and can achieve adequate detection accuracy and real-time performance on the embedded device.

Weed Detection in Peanut Fields Based on Machine Vision

The accurate identification of weeds in peanut fields can significantly reduce the use of herbicides in the weed control process. To address the identification difficulties caused by the cross-growth of peanuts and weeds and by the variety of weed species, this paper proposes a weed identification model named EM-YOLOv4-Tiny incorporating multiscale detection and attention mechanisms based on YOLOv4-Tiny. Firstly, an Efficient Channel Attention (ECA) module is added to the Feature Pyramid Network (FPN) of YOLOv4-Tiny to improve the recognition of small target weeds by using the detailed information of shallow features. Secondly, the soft Non-Maximum Suppression (soft-NMS) is used in the output prediction layer to filter the best prediction frames to avoid the problem of missed weed detection caused by overlapping anchor frames. Finally, the Complete Intersection over Union (CIoU) loss is used to replace the original Intersection over Union (IoU) loss so that the model can reach the convergence state faster. The experimental results show that the EM-YOLOv4-Tiny network is 28.7 M in size and takes 10.4 ms to detect a single image, which meets the requirement of real-time weed detection. Meanwhile, the mAP on the test dataset reached 94.54%, which is 6.83%, 4.78%, 6.76%, 4.84%, and 9.64% higher compared with YOLOv4-Tiny, YOLOv4, YOLOv5s, Swin-Transformer, and Faster-RCNN, respectively. The method has much reference value for solving the problem of fast and accurate weed identification in peanut fields.

Automatic Extraction of Power Lines from Aerial Images of Unmanned Aerial Vehicles.

Automatic power line extraction from aerial images of unmanned aerial vehicles is one of the key technologies of power line inspection. However, the faint power line targets and complex image backgrounds make the extraction of power lines a greater challenge. In this paper, a new power line extraction method is proposed, which has two innovative points. Innovation point one, based on the introduction of the Mask RCNN network algorithm, proposes a block extraction strategy to realize the preliminary extraction of power lines with the idea of “part first and then the whole”. This strategy globally reduces the anchor frame size, increases the proportion of power lines in the feature map, and reduces the accuracy degradation caused by the original negative anchor frames being misclassified as positive anchor frames. Innovation point two, the proposed connected domain group fitting algorithm solves the problem of broken and mis-extracted power lines even after the initial extraction and solves the problem of incomplete extraction of power lines by background texture interference. Through experiments on 60 images covering different complex image backgrounds, the performance of the proposed method far exceeds that of commonly used methods such as LSD, Yolact++, and Mask RCNN. DSCPL, TPR, precision, and accuracy are as high as 73.95, 81.75, 69.28, and 99.15, respectively, while FDR is only 30.72. The experimental results show that the proposed algorithm has good performance and can accomplish the task of power line extraction under complex image backgrounds. The algorithm in this paper solves the main problems of power line extraction and proves the feasibility of the algorithm in other scenarios. In the future, the dataset will be expanded to improve the performance of the algorithm in different scenarios.

Detection and Counting Method of Pigs Based on YOLOV5_Plus: A Combination of YOLOV5 and Attention Mechanism

Information-based pig detection and counting is the trend in smart animal husbandry development. Cameras can efficiently collect farm information and combine it with artificial intelligence technology to assist breeders in real-time monitoring and analysis of farming. In order to improve the speed and accuracy of pig detection and counting, an advanced improved YOLO_v5 method for pig detection and counting based on the attention mechanism is proposed. The model is named as YOLOV5_Plus. This article utilizes a series of data augmentation methods, including translation, color augmentation, rescaling, and mosaic. The proposed model performs feature extraction on the original image with a backbone network, detects pigs of different sizes with three detection heads, and counts the detected anchor frames. Different versions of YOLOV5 are compared, and YOLOV5x is selected as the baseline model for the best performance. Attention modules are smartly combined with the model so that the model can better handle overlapping and misidentification. YOLOV5_Plus can achieve an accuracy of 0.989, a recall of 0.996, mAP@.50 of 0.994, and mAP@.50:.95 of 0.796, which outperforms all competing models. The inference time per image during detection is only 24.1 ms. YOLOV5_Plus model achieves real-time pig number and location detection, which is meaningful for promoting smart animal husbandry and saving labor costs in farming enterprises.

Using Pruning-Based YOLOv3 Deep Learning Algorithm for Accurate Detection of Sheep Face.

Simple SummaryThe identification of individual animals is an important step in the history of precision breeding. It has a great role in both breeding and genetic management. The continuous development of computer vision and deep learning technologies provides new possibilities for the establishment of accurate breeding models. This helps to achieve high productivity and precise management in precision agriculture. Here, we demonstrate that sheep faces can be recognized based on the YOLOv3 target detection network. A model compression method based on K-means clustering algorithm and combined with channel pruning and layer pruning is applied to individual sheep identification. In addition, the results show that the proposed non-contact sheep face recognition method can identify sheep quickly and accurately.Accurate identification of sheep is important for achieving precise animal management and welfare farming in large farms. In this study, a sheep face detection method based on YOLOv3 model pruning is proposed, abbreviated as YOLOv3-P in the text. The method is used to identify sheep in pastures, reduce stress and achieve welfare farming. Specifically, in this study, we chose to collect Sunit sheep face images from a certain pasture in Xilin Gol League Sunit Right Banner, Inner Mongolia, and used YOLOv3, YOLOv4, Faster R-CNN, SSD and other classical target recognition algorithms to train and compare the recognition results, respectively. Ultimately, the choice was made to optimize YOLOv3. The mAP was increased from 95.3% to 96.4% by clustering the anchor frames in YOLOv3 using the sheep face dataset. The mAP of the compressed model was also increased from 96.4% to 97.2%. The model size was also reduced to 1/4 times the size of the original model. In addition, we restructured the original dataset and performed a 10-fold cross-validation experiment with a value of 96.84% for mAP. The results show that clustering the anchor boxes and compressing the model using this dataset is an effective method for identifying sheep. The method is characterized by low memory requirement, high-recognition accuracy and fast recognition speed, which can accurately identify sheep and has important applications in precision animal management and welfare farming.

Improved Faster R-CNN Based Surface Defect Detection Algorithm for Plates

Defect recognition plays an important part of panel inspection, and most of the current manual inspection methods are used, but the recognition efficiency and recognition accuracy are low. The Fast-Convolutional Neural Network (Faster R-CNN) algorithm is improved, and a surface defect detection algorithm based on the improved Faster R-CNN is proposed. Firstly, the algorithm improves the bilateral filtering algorithm to smooth the image texture background. Subsequently, a feature pyramid network with a shape-variable convolutional ResNet50 network can be applied to acquire defect semantic feature maps to improve the network's ability to express the features of multiscale defects while solving the difficulty problem of many types of defects and variable shapes. To obtain more accurate defect localization information, the algorithm in this paper uses the Region of Interest Align (ROI Align) algorithm instead of the crude Region of Interest Pooling (ROI Pooling) algorithm. Then, an improved attention region recommendation network is used to improve the focus of the model on plate defects and suppress the features of complex background. Finally, a K-means algorithm is added to cluster the defect data to derive anchor frames that are better adapted to the plate defects. In this paper, a dataset containing 3216 images of surface defects of plate metal is made by acquiring surface defect images from the production site of the plate metal factory, which mainly include various defect types. This dataset is used to train and test the algorithm model of this paper, and the results of detection accuracy and detection speed are compared with those of other algorithms, which prove that the algorithm of this paper can achieve real-time detection of plate defects with high detection accuracy.

A high-precision forest fire smoke detection approach based on ARGNet

The occurrence of forest fires can lead to ecological damage, property loss, and human casualties. Current forest fire smoke detection methods do not sufficiently consider the characteristics of smoke with high transparency and no clear edges and have low detection accuracy, which cannot meet the needs of complex aerial forest fire smoke detection tasks. In this paper, we propose Adjacent layer composite network based on a recursive feature pyramid with deconvolution and dilated convolution and global optimal nonmaximum suppression (ARGNet) for high-accuracy detection of forest fire smoke. First, the Adjacent layer composite network is proposed to enhance the extraction of smoke features with high transparency and no clear edges, and SoftPool in it is used to retain more feature information of smoke. Then, a recursive feature pyramid with deconvolution and dilated convolution (RDDFPN) is proposed to fuse shallow visual features and deep semantic information in the channel dimension to improve the accuracy of long-range aerial smoke detection. Finally, global optimal nonmaximum suppression (GO-NMS) sets the objective function to globally optimize the selection of anchor frames to adapt to the aerial photography of multiple smoke locations in forest fire scenes. The experimental results show that the ARGNet parametric number on the UAV-IoT platform is as low as 53.48 M, mAP reaches 79.03%, mAP50 reaches 90.26%, mAP75 reaches 82.35%, FPS reaches 122.5, and GFLOPs reaches 55.78. Compared with other mainstream methods, it has the advantages of real-time detection and high accuracy.

Research on Vehicle Lane Change Warning Method Based on Deep Learning Image Processing.

In order to improve vehicle driving safety in a low-cost manner, we used a monocular camera to study a lane-changing warning algorithm for highway vehicles based on deep learning image processing technology. We improved the mask region-based convolutional neural network for vehicle target detection. Suitable anchor frame ratios were obtained by means of K-means++ method clustering for 66,389 vehicle targets with the width/height ratio, which is one more set of anchor frames than the original setting, so as to ensure that the generation accuracy of candidate frames can be improved without sacrificing more network performance. Using the vehicle target annotation set, we trained the vehicle targets. Through the analysis of indicators for mean average precision, a new set of anchor frames was added to improve the accuracy of vehicle target detection. Based on the improved vehicle detection network and an end-to-end lane detection network in series, we proposed an algorithm for the detection of highway vehicle lane-changing behavior with the first-person perspective by summing the inter-frame change rates in the vehicle lane-changing data pool. After the identification and verification of the marked lane-changing picture sequences, a lane-changing detection accuracy rate of 94.5% was achieved.

Dynamic Selection of Optional Feature for Object Detection

To obtain the most intuitive pedestrian target detection results and avoid the impact of motion pose uncertainty on real-time detection, a pedestrian target detection system based on a convolutional neural network was designed. Dynamic Selection of Optional Feature (DSOF) module and a center branch were proposed in this paper, and the target was detected by an anchor-free method. Although almost all the most advanced target detectors use pre-defined anchor boxes to run through the possible positions, scales, and aspect ratios of search targets, their effectualness, and generalization ability are also limited by the anchor boxes. Most anchor-based detectors use heuristically guided anchor frames. Such a design is difficult to detect objects of different types and sizes, especially objects with highly overlapping boundaries. To solve this problem, the DSOF module is proposed in this paper, which selects for each instance the most appropriate feature layer through automatic feature selection. After using multi-level prediction, stacks of low-grade prediction bounding boxes will be generated far away from the target center. To eliminate these low-grade detections, we introduce a new center branch to predict the deviation of a pixel from its corresponding bounding box. This score is used to reduce the weight of the low-grade detection bounding box and merge the detection sequences into the Non-Maximum Suppression (NMS).

Research on traffic sign recognition method based on multi-scale convolution neural network

In order to accurately identify the traffic sign information under different road conditions, an improved deep learning method based on Faster RCNN model is proposed. Firstly, a multi-channel parallel full convolution neural network is designed to extract the color, shape and texture features of traffic signs in the original image. The multi-channel feature layers are fused to get the final feature map, and the adaptability of the model in various environment and weather conditions is enhanced by the image preprocessing. At the same time, the fusion features of deep and shallow feature layer are added into the feature extraction network, and the detailed texture information of shallow feature layer and semantic information of deep feature layer are retained, and the final feature layer can adapt to multi-scale change of traffic sign recognition. Secondly, the prior knowledge of traffic signs is used to detect and locate the target before the original RPN candidate region is generated. A more reasonable method for generating feature points and candidate anchor frames for traffic sign recognition is proposed. Based on the prior knowledge statistics of traffic sign size and proportion results, a target candidate frame suitable for traffic sign recognition is designed, a large number of redundant and negative correlation candidate frames is reduced, the detection accuracy and reduces the detection time is improved; secondly, the multi-scale candidate frame generation method for the deep and shallow feature layer is added to enhance the multi-scale target recognition ability and further strengthen the multi-scale target recognition ability Finally, this paper uses the international general traffic sign specification data set GTSRB/GTSDB and domestic traffic sign data set tt100k to verify the recognition ability of the model.

Assessing Label Frames and Emotional Primes in the Context of Animal Rearing—Response of an Explorative fNIRS Study

Against the background of rising societal interest for sustainable food and nutrition choices, food labels have gained importance in providing important information to consumers. However, little is known about how the differences between quality frames in labels are evaluated and how priming might serve as an anchor for label perception. This study aims to observe the neural reaction of this in the context of differently framed food labels for products of animal origin, claiming the presence or absence of an additional quality aspect and under the impulse of emotional priming. In an explorative setup, we measured the neural prefrontal cortex activity of 26 participants with the neuroimaging technology fNIRS. An idyllic prime and a prime related to a label claiming an additional product quality led to increased neural activity in the OFC and dlPFC. Shedding light on what elements are of importance to identify products that meet consumers’ requirements in terms of quality aspects, this could indicate that the prime stressed the meaning of the label. This strengthens the argument to positively phrase and anchor frames regarding quality attributions as opposed to negative declarations. It further demonstrates the ability of fNIRS to capture processing through labels and primes in the context of consumer behavior.

Automatic fabric defect detection using a wide-and-light network

Automatic fabric defect detection systems improve the quality of textile production across the industry. To make these automatic systems accessible to smaller businesses, one potential solution is to use limited memory capacity chips that can be used with hardware platforms with limited resources. That is to say, the fabric defect detection algorithm must ensure high detection accuracy while maintaining a low computational cost. Therefore, we propose a wide-and-light network structure based on Faster R-CNN for detecting common fabric defects. We enhance the feature extraction capability of the feature extraction network by designing a dilated convolution module. In a dilated convolution module, a multi-scale convolution kernel is used to adapt to defects of different sizes. Dilated convolutions can increase receptive fields without increasing the number of parameters used. Therefore, we replace a subset of ordinary convolutions with dilated convolutions to learn target features and use convolution kernel decomposition and bottleneck methods to simplify the feature extraction networks. Then, high-level semantic features are fused with bottom-level detail features (via skip-connection) to obtain multi-scale fusion features. Finally, a series of anchor frames (of different sizes) is designed to suit multi-scale fabric defect detection. Experiments show that compared with various mainstream target detection algorithms, our proposed algorithm can improve the accuracy of fabric defect detection and reduce the size of the model.

Research on Taproots Identification Technology in Panax notoginseng Quality Intelligent Management System.

In the Panax notoginseng quality intelligent management system, the big roots and fibrous roots cannot be cut automatically because the machine cannot distinguish the taproot, big roots, and fibrous roots of Panax notoginseng, resulting in the automatic cutting mechanism unable to obtain the control trajectory coordinate reference of the tool feed. To solve this problem, this paper proposes a visual optimal network model detection method, which uses the image detection method of marking anchor frames to improve the detection accuracy. A variety of deep learning network models are modified by the TensorFlow framework, and the best training model is optimized by comparing the results of training, testing, and verification data. This model is used to automatically identify the taproots and provide the control trajectory coordinate reference for the actuator that cuts big roots and fibrous roots automatically. The experimental results show that the optimal network model studied in this paper is effective and accurate in identifying the taproots of Panax notoginseng.

High precision detection algorithm based on improved RetinaNet for defect recognition of transmission lines

The Unmanned Aerial Vehicle (UAV) inspection mode has been gradually implemented in the power system. The UAV inspection image is checked by the target detection technology, but there is no high-precision target detection algorithm as the technical support. In this regard, this paper proposes a target detection algorithm based on the improved RetinaNet which is suitable for transmission lines defect detection. In this algorithm, the shortcomings of the RetinaNet anchor frame extraction mechanism based on Apriori are corrected. At the same time, the number and size of anchor frames are redesigned by using the improved K-means + + algorithm, so that the anchor frame of the improved algorithm gets the highest average IoU (Intersection over Union) value, which matches the actual size of the transmission line defects. Then, in RetinaNet, the feature pyramid network based on DenseNet is built as the backbone network to improve the model accuracy and make the model lighter. The improved model is trained and tested by using the data set of transmission line defects for validation. The results show that the proposed method has advantages and effectiveness in the detection of transmission line defects, and meets the requirements of intelligent inspection in terms of accuracy.