It is estimated that 50% of the overhead line maintenance costs are related to insulators diagnostics, replacement and repair. Insulator failures cause about 70% of downtime in line operation [1]. This picture shows the importance of diagnostics and early detection of insulator failures. Diagnostic tests of insulators serve three main tasks: identification of damaged insulators and those posing a high risk of failure, assessment of the degree of ageing degradation of the properties of insulators, and detection of poorly made or incorrectly selected insulators. Insulator diagnostics generally can be divided into two different types: tests in laboratory conditions and inspections of overhead lines [24]. Condition inspections of insulators during operation are essential because today’s electricity demand puts enormous pressure on electricity distributors to minimise the downtime of the overhead lines. Assessing the condition of power lines is the basic activity that directly impacts reducing the number of failures in distribution networks. From the power system security viewpoint, high and extra-high voltage lines have strategic importance.

Collecting diagnostic information about power lines can be divided into:

• actions performed from the ground by human walking teams and off-road vehicles;
• aerial operations with the use of helicopters and UAVs [15,25,26];
• activities performed by devices mounted directly or close to power line elements, for example, various types of monitoring systems and diagnostic robots [27,28,29,30].

The dynamic development of various types of vision systems used for the observation of public spaces and the protection of various types of industrial facilities allows for their simple application in the electrical power industry in combination with computer image analysis techniques. Recognising the condition of electricity infrastructure objects in digital images creates a wide range of potential applications. These may include [29,31,32,33]: precise measurements of the position of line structural elements, monitoring to support early warning systems, condition monitoring, and fault localisation in high-voltage networks, identification of line hazards arising from the immediate environment. The data collected during a ground inspection are usually images obtained using various cameras. Their interpretation is complicated, as they are usually taken from large angles and considerable distances. This is due to the specificity of overhead line construction—the most critical components, for example, wires, insulators, and connectors, are suspended over a considerable height, as it is crucial to isolate the conducting wires from the ground. The data obtained during helicopter or aeroplane overflights allows a more accurate analysis of the essential elements of the line, as such images are usually taken from observation points suspended directly above the line. On the other hand, using UAVs allows line elements to be analysed from virtually any angle and relatively short distances. The use of UAVs means that the inspection of an overhead line can be more efficient and accurate, as the drone can fly closer to the line than a helicopter or foot patrol [34].

In addition, inspections carried out by unmanned aerial vehicles allow for a significant reduction in the cost of the entire process and shorten the time needed to perform the diagnostics of the overhead power line. Currently, there are more and more different flying platforms offering different possibilities of imaging objects (i.e., resolutions and types of imaging). An interesting option is also the possibility of using a swarm of drones, which allows for the simultaneous imaging of overhead lines by several vehicles from different perspectives. The problem is the massive amount of data received, which must be analysed manually afterwards [31]; therefore, it is necessary to create new efficient methods of their processing.

The paper [35] presents a simple method for detecting insulators in aerial photographs by binarising the image and applying morphological operations. However, the detection is limited to tempered glass insulators only, and the criteria for selecting the adaptive threshold for different lighting conditions are not included in this paper. Other work [36,37] has suggested using colour features extraction to detect insulators. In work [38], colour-based segmentation was used to separate the insulator from the background.

In publications [39,40], edge-based feature extractors have been used to detect porcelain insulators from images taken with drones and cleaning robots. The methods presented performed poorly on images where the background was not uniform. A publication by Zhao et al. [41] proposed an insulator lattice model by grouping similar appearance glass and porcelain insulator components together and then performing a network search using the Markov Random Field (MRF) algorithm. The extracted data are then combined with spatial contextual information to localise multiple insulators quickly. The proposed method works stably on a complex background, but its performance is only guaranteed when a group of insulators appear together in an image, which significantly limits its application.

Liao and An in [42] proposed a robust Multiscale and Multi-Feature (MSMF) descriptor based on local features. On the other hand, in [43], Haar features and the AdaBoost classifier were used to detect the insulator. In this paper, a synthesised 3D model of the insulator was used to train the classifier. It showed a significant improvement in detection accuracy. However, both methods [42,43] only detect insulators in images taken from a long distance with low resolution, which is unsuitable for further defect analysis.

Li et al. in [44] used Vertical Profile Projection curves (VPP) as features to determine the shape of insulators and a Support Vector Machines (SVM) classifier to detect them. Wang et al. [34] proposed a Gabor feature detector and SVM classifier for insulator detection. The methods proposed in [34,39] are based on a repeating pattern on the insulator. However, insulators are only well observed when a picture is taken perpendicularly to a power line. As a result, this method’s photo of the insulator taken at any angle is not usable. Li et al. [45] proposed using a local and global relevance map to segment insulators. However, their method only works when the texture and intensity of the background and foreground areas are clear. Such a condition usually occurs only when aerial photographs of insulators are taken from closer distances and using appropriate camera optics.

The method developed in [46] based on Speeded Up Robust Features (SURF) and Intuitionistic Fuzzy Set (IFS) algorithms allows localising in aerial photographs without using pattern and segmentation. The first step searches for crucial features using the SURF algorithm. Then the obtained points are divided into a certain number of classes using the IFS algorithm based on the correlation coefficient. If the correlation between the obtained sets is more significant than the set value, then both classes can be treated as sets of the same class. The insulator is identified based on characteristic shape factors values such as slenderness or duty ratios. Another approach is to locate an insulator based on colour, an example being the research published in [47]. Methods based on the SIFT and SURF algorithms can locate an object accurately; however, their application has some limitations. Depending on the complexity of the background, they generate large numbers of significant (local features), which translates into increased computational costs. In the case of aerial photographs, insulators are most often located against very different backgrounds. A previously created pattern is also most often required to locate the feature.

Oberweger et al. [48] presented a novel approach for detecting insulators in aerial photographs. They based their algorithm on discriminative training of local gradient-based feature descriptors and a voting scheme based on the Random Sample Consensus (RANSAC) algorithm. However, their algorithm does not allow the detection of multiple insulators in a single image.

Another different approach to trying to locate an insulator but using pattern matching is the approach presented in [49]. In this paper, the author segments the image into specific classes using Statistical Region Merging (SRM) and then converts the image to greyscale for histogram analysis. The histogram at this stage represents the individual objects in the image. Insulator identification is based on pattern matching using the correlation method. However, this method cannot cope with irregularities in the insulator structure and is sensitive to noise in the image.

Jabid and Uddin [50] used a classical detection method based on a sliding window allowing the detection of Local Directional Pattern (LDP) features and an SVM classifier. Their method not only needs to scale the input image to multiple sizes but also rotates the input image in multiple orientations to account for changes in size and rotation, which significantly slows down the detection process.

In [51], the authors used a cascaded CNN architecture based on Reverse Polish Notation (RPNs). This work combined VGG Neural Networks and ResNet networks, but the limitation is the insufficient speed for real-time operation. In the paper [35], a convolutional neural network is used for insulator feature extraction and classification, and OTSU-based segmentation is applied in the next step. In [52], a deep learning algorithm based on feature detection is proposed. Region Proposal Network (RPN) is used to generate region proposals, and a Fully Convolutional Network (FCN) is used to obtain object maps.

The above methods were tested on images taken with aircraft and helicopters. As these vehicles have several disadvantages, for example, high operating price, complicated operation, and susceptibility to bad weather conditions (e.g., strong wind), insulator detection systems based on ground vehicles have also been proposed [40,50,53,54] Li et al., in their paper [40], used an improved MPEG-7 edge histogram descriptor to detect insulators on frames from videos taken from the ground. The publication [54] used a detector based on a wavelet transform and an SVM classifier to detect insulators. Another publication by the same authors [53] extracted features using a wavelet transform and then used a hidden Markov model to classify damaged insulators.