Convolutional neural networks (CNNs) are deep neural networks that can be trained on large databases and show outstanding performance on object classification, segmentation, image denoising etc. In the past few years, several image denoising techniques have been developed to improve the quality of an image. The CNN based image denoising models have shown improvement in denoising performance as compared to non-CNN methods like block-matching and three-dimensional (3D) filtering, contemporary wavelet and Markov random field approaches etc. which had remained state-of-the-art for years. This study provides a comprehensive study of state-of-the-art image denoising methods using CNN. The literature associated with different CNNs used for image restoration like residual learning based models (DnCNN-S, DnCNN-B, IDCNN), non-locality reinforced (NN3D), fast and flexible network (FFDNet), deep shrinkage CNN (SCNN), a model for mixed noise reduction, denoising prior driven network (PDNN) are reviewed. DnCNN-S and PDNN remove Gaussian noise of fixed level, whereas DnCNN-B, IDCNN, NN3D and SCNN are used for blind Gaussian denoising. FFDNet is used for spatially variant Gaussian noise. The performance of these CNN models is analysed on BSD-68 and Set-12 datasets. PDNN shows the best result in terms of PSNR for both BSD-68 and Set-12 datasets.

Research in human activity recognition (HAR) has seen tremendous growth and continuously receiving attention from both the Computer Vision and the Image Processing communities. Due to the existence of numerous publications in this field, undoubtedly, there have been a number of review papers on this subject that categorise these techniques. Many of the recent works have started to tackle more challenging problems and these proposed techniques are addressing more realistic real-world scenarios. Conspicuously, an updated survey that covers these methods is timely due. To simplify the categorisation, this study takes a two-layer hierarchical approach. At the top level, the categorisation is based on the basic process flow of HAR, i.e. input data-type, features-type, descriptor-type, and classifier-type. At the second layer, each of these components is further subcategorised based on the diversity of the proposed methods. Finally, a remark on the coming popularity of deep learning approach in this field is also given.

Image clustering is considered amongst the most important tasks in medical image analysis and it is regularly required as a starter and vital stage in the computer-aided medical image process. In brain magnetic resonance imaging (MRI) analysis, image clustering is regularly used for estimating and visualising the brain anatomical structures, to detect pathological regions and to guide surgical procedures. This study presents a new method for MRI brain images clustering based on the farthest point first algorithm and fuzzy clustering techniques without using any a priori information about the clusters number. The algorithm has been approved against both simulated and clinical magnetic resonance images and it has been compared with the fourth clustered algorithms. Results demonstrate that the proposed algorithm has given reasonable segmentation of white matter, grey matter and cerebrospinal fluid from MRI data, which is superior in preserving image details and segmentation accuracy compared with the other four algorithms giving more than 91% in Jaccard similarity.

Glaucoma is a critical and irreversible neurodegenerative eye disorder caused by damaging optical nerve head due to increased intra-ocular pressure within the eye. Detection of glaucoma is a critical job for ophthalmologists. This study presents a novel and more accurate method for automated glaucoma detection using quasi-bivariate variational mode decomposition (QB-VMD) from digital fundus images. In total, 505 fundus images are decomposed using QB-VMD method which gives band limited sub-band images (SBIs) centred around a particular frequency. These SBIs are smooth and free from mode mixing problems. The glaucoma detection accuracy depends on the most useful features as it captured appropriate information. Seventy features are extracted from QB-VMD SBIs. Extracted features are normalised and selected using ReliefF method. Selected features are then fed to singular value decomposition to reduce their dimensionality. Finally, the reduced features are classified using least square support vector machine classifier. The obtained glaucoma detection accuracies are 85.94 and 86.13% using three- and ten-fold cross validation, respectively. Obtained results are better than the existing. It may become a suitable method for ophthalmologists to examine eye disease more accurately using fundus images.

This study presents a fast and robust ellipse detector based on edge following method. The detector first extracts segments using an edge predictor based on curvature analysis. Then, line segments are generated based on length condition other than least-squares approximation. After that, potential ellipses are detected based on edge curvature and convexity. In addition, a re-find contours detection method is introduced to improve the accuracy by searching edge points in the missing part of the ellipse. The performance of the detector has been tested on different datasets containing both synthetic and real images with three other algorithms based on the edge following method. Experimental results indicate that the proposed method always has the fastest execution time. Besides, it advances the state of the art in accuracy in most cases. Generally speaking, it is a fast, robust and effective ellipse detector for real-time applications.

Telemedicine is getting a lot of popularity. It is because of the way that the medical images are sent over the Internet to share the restorative state of a patient, making the analysis simple for the medical staff. There is, however, a good possibility that the contents of a medical image may be altered due to the openness of the Internet. The watermarking techniques are utilised to keep the integrity of medical images intact. These techniques, however, need some additional information (watermark) to be embedded into the medical images, which may deteriorate their visual quality, leading to false diagnosis. In this study, a non-intrusive procedure for verification of the medical images is proposed in which the polar cosine transform is used for feature extraction and the PatchMatch algorithm to locate the forged regions. Experimentally, the authors demonstrate that the proposed technique is robust against the noise, blurring, scaling, rotation and JPEG compression.

Objective evaluation of a subjective image quality assessment plays a significant role in the various image processing applications, such as compression, interpolation and noise reduction. The subjective image quality assessment does not only depend on some objective measurable artefacts, but also on image content and kind of distortions. Thus, a multi-parameter prediction of the objective image quality assessment is proposed in this study. The prediction parameters are found minimising the mean square error related to the known subjective image quality measure (DMOS). This approach includes mostly used image quality metrics (peak signal-to-noise ratio, multi-scale structural similarity image measure, feature similarity image measure, video quality measure) and two-dimensional image quality metrics (2D IQM). The proposed multi-parameter prediction has been verified on the test image database (LIVE) for compression, noise and blur distortions with available subjective image quality measures (DMOS). More reliable estimations are obtained using multi-parameter prediction instead of only one measure. The best results are reached when an image content indicator is combined with the 2D IQM measure separately for different kinds of distortions.

In many applications in order to recognise the relationship between user and computer, the position at which the user looks should be detected. To this end, a salient object should be extracted that is attracted to the attention of the viewer. In this study, a new method is proposed to extract the object saliency map, which is based on learning automata and sparse algorithms. In the proposed method, after decomposition of an image to its superpixels, eight features (namely three features in red–green–blue colour space, coalition, central bias, rotation feature, brightness, and colour difference) are extracted. Then the extracted features are normalised to zero mean and unit variance. In this study, K-means singular-value decomposition is used to integrate the extracted features. The performance of the proposed method is compared with that of 20 other methods by applying four new databases, including MSRA-100, ECSSD, MSRA-10K, and Pascal-S. The obtained results show that the proposed method has a better performance compared to the other methods with regard to the prediction of the salient object.

Images captured under low-light conditions are often of low visibility. To improve visualisation, a novel low light image enhancement method is presented based on the non-uniform illumination prior model. First, the k-means method is used to process the value channel in the hue-saturation-value (HSV) colour space after space conversion of the input image. Then, the initial illumination of segmented scenes is estimated by an improved maximum red–green–blue method. Next, an illumination preservation method is presented to maintain the naturalness of the enhanced image. Furthermore, the non-uniform illumination prior model is proposed to enhance the textural details in the enhanced image. Fast Fourier transformation is used to accelerate the optimisation. Since an adaptive weight is assigned, the proposed method can preserve the edges and textures at the bright and edge areas. Experimental analysis shows that the results using the proposed method have less noise, better illumination, improved contrast, and satisfactory naturalness. In addition, the proposed method can provide better quality images in terms of subjective and objective assessments.

In this study, the authors propose a novel three-dimensional (3D) convolutional neural network for shape reconstruction via a trilateral convolutional neural network (Tri-CNN) from a single depth view. The proposed approach produces a 3D voxel representation of an object, derived from a partial object surface in a single depth image. The proposed Tri-CNN combines three dilated convolutions in 3D to expand the convolutional receptive field more efficiently to learn shape reconstructions. To evaluate the proposed Tri-CNN in terms of reconstruction performance, the publicly available ShapeNet and Big Data for Grasp Planning data sets are utilised. The reconstruction performance was evaluated against four conventional deep learning approaches: namely, fully connected convolutional neural network, baseline CNN, autoencoder CNN, and a generative adversarial reconstruction network. The proposed experimental results show that Tri-CNN produces superior reconstruction results in terms of intersection over union values and Brier scores with significantly less number of model parameters and memory.

In this study, a method for fusing of visible (or standard RGB) with near-infrared (NIR) image pair for enhancing a hazy image is proposed. Better image enhancement in terms of contrast, sharpness, increased perception is realised by combining the components from both the spectra. While there are a number of applications that use NIR images, very few combine RGB and NIR information of the same scene taken from two separate imaging devices. Although NIR images are greyscaled in nature, they have intrinsic properties desirable in both colour and visible imagery of a number of scenes. Instances are increased the contrast between sky and clouds; shadowed and non-shadowed regions; and increased optical depth. These features are in most cases missed by a standard RGB camera and therefore, a fusion of such RGB–NIR image pairs is highly beneficial. The scheme for fusing these image pairs is realised by using a novel, fuzzy clustering algorithm along with wavelet transform so that not only contrast and sharpness are enhanced, but also the chromaticity of the original RGB image is retained. The proposed technique is well demonstrated for various outdoor scenes, and better results are obtained when compared against some of the state-of-the-art techniques proposed recently.

The current study presents a new 3D mesh deformation process using multi-mother wavelet neural network architecture, which relies on genetic algorithm and multiresolution analysis. Classic forming algorithms begin with a predetermined network architecture that may be insufficient or too complicated. In addition, the solving of wavelet neural network training problems is described by their perceived inability to escape local optima. The main objective of the authors' proposed approach is that it prevents both the insufficiency and local minima by integrating the genetic algorithm; their wavelet network is used as an approximation tool to align the features of mesh to have efficient deformation processes. Also, such meshes are especially expensive to transmit and are awkward to deform. For this reason, they propose to use multiresolution analysis to decompose a surface geometry into several levels of detail in order to work only with the approximation coefficient at a chosen decomposition level. Hierarchical triangle mesh representations provide access to a triangle mesh at the desired resolution without omitting any information. The experimental results showed the validity of the generalisation ability and the efficiency of their suggested multi-mother wavelet network architecture based on genetic algorithm and multiresolution analysis for 3D mesh modelling and deformation.

It is known that parameters such as the feed rate and the spindle speed affect the hole quality during the drilling of aluminium and its alloys. In particular, deflection occurs as a result of the increase or decrease of the reverse forces acting on the tool as a result of changing the parameter values. The tool deflection causes deviations in the hole geometry. This requires the initial detection of the deflection on the tool and then the most appropriate updating of the drilling parameters. At the present time, force-based estimation and inductive or laser sensor detection methods are used for the detection of tool deflection. These methods are useless because they require expensive measurement systems and continuous fine-tuning. This study aimed to calculate the tool deflection that occurs during the drilling of AL 7075 material using an image processing technique. In the experiments using different drilling parameters, the tool deflection was calculated and the effects of the parameters on tool deflection were investigated. As a result, it is shown that the tool deflection can be detected quickly and simply by image processing. In addition, the effects of the processing parameters on the tool deflection are discussed.

Image upscaling is needed in many areas. There are two types of methods: methods based on a simple hypothesis and methods based on machine learning. Most of the machine learning-based methods have disadvantages: no support is provided for a variety of upscaling factors, a training process with a high time cost is required, and a large amount of storage space and high-end equipment are required. To avoid the disadvantages of machine learning, upscaling images with a simple hypothesis is a promising strategy but simple hypothesis always produces jaggy artifacts. The authors propose a new method to remove these jagged artifacts. They consider an edge in an image as a deformed curve. Removing jagged artefacts is considered equivalent to shortening the full arc length of a curve. By optimising the regularization model, the severity of the artifacts decreases as the number of iterations increases. They compare nine existing methods on the Set5, Set14, and Urban100 datasets. Without using any external data, the proposed algorithm has high visual quality, has few jagged artefacts and performs similarly to very recent state-of-the-art deep convolutional network-based approaches. Compared to other methods without external data, the proposed algorithm balances the quality and time cost well.

In this study, an image representation method based on multi-scale microstructural binary pattern extraction is proposed, which uses zero-mean microstructural pattern binarisation. This method can express all kinds of important pattern structures that may appear in the image. By using the dominant binary pattern learning model, the dominant feature pattern sets adapted to different datasets can be obtained, which have good performance in the aspects of feature robustness, recognition, and representation ability. This method can greatly reduce the dimension of feature coding and improve the speed of the algorithm. The experimental results show that this method has strong recognition ability and robustness, is superior to the traditional local binary pattern and grey image micorstructure maximum response pattern methods, and has a competitive performance compared with the results of many latest algorithms.

A unified framework which provides a higher security level to e-passports is proposed. This framework integrates face, iris and fingerprint images. It involves three layers of security: the first layer maps a biometric image to another biometric image which is called biostego image. Three mapping schemes are proposed: the first scheme maps single biometric image to single biostego image, the second scheme maps dual biometric images to single biostego image, the third scheme divides the biometric image into sections and maps each section to different biostego image. A mapping function maps the intensity value of each pixel in the biometric image to pixels with same intensity in the biostego image. A representative pixel is randomly selected from the set of pixels, and its coordinates are recorded in the location map of the biometric image. In the second layer, the location map is encoded using fingerprint fuzzy vault. In the third layer, the encoded location map is hidden in the biostego image using steganography technique. The biostego image which contains the encoded location map is stored in the e-passport's memory. Keeping the mapping scheme secret and by using the fingerprints fuzzy vault to encrypt location map, the proposed approach provides higher level of protection against fraud.

Cross-modal hashing (CMH) has received widespread attention due to high retrieval efficiency, which plays an extremely important role in cross-modal retrieval. Recently, many CMH methods have been proposed to establish the semantic connection of different modalities. However, most of these methods only use a simple quantisation strategy, resulting in large quantisation error, and inferior hash codes. To address this issue, in this study, the authors propose a novel self-taught CMH (STCMH) to minimise the semantic encoding loss. In particular, the common semantic representations across different modalities are first learnt based on collective matrix factorisation. Then, the quantisation procedure based on orthogonal transformation is integrated to encode the semantic representations into discriminative binary codes. Moreover, similarity preservation is imposed to further boost the discriminative power. Finally, hashing functions learning is formulated as a binary classification problem by self-taught scheme. Experimental results on three public datasets demonstrate that STCMH significantly outperforms most state-of-the-art CMH methods.

The morphological structure of retinal vessels plays an important role in analysing and diagnosing fundus disease. In this study, an unsupervised automatic segmentation method for retinal blood vessels with corrected morphological transformation and fractal dimension is proposed. To enhance the contrast between retinal vessels and background in a fundus image, the morphological operator with linear structural elements is used; to remove the lesion and its light reflection, a compensation method is proposed; to extract the objects from a grey image, the multi-threshold approach is applied; to recognise the blood vessels and noise from the fundus image, a new method based on fractal dimension is presented. The new approach is tested in detail on three public databases STARE, DRIVE and HRF. Experimental results show that the segmentation algorithm is better than other existing unsupervised automatic segmentation algorithms, and the new approach is robust.

An automatic and novel approach for acute lymphoblastic leukaemia classification is proposed. The proposed scheme is based on pre-processing and segmentation of white blood cell nuclei using expectation maximisation algorithm, feature extraction, feature selection using principal component analysis and classification using sparse representation. The accuracy of the proposed scheme significantly outperforms the existing schemes in terms of acute lymphoblastic leukaemia classification.

This study proposes a multi-exposure image fusion (MEF) technique that takes multi-exposure input images and produces a high-quality output image without any artefacts. The proposed technique consists of four steps. In the first step, three quality measures (contrast, saturation, and well exposedness) are measured. Secondly, the colour dissimilarity approach is used to detect moving objects. Thirdly, the authors calculate the weight maps using three quality measures (contrast, saturation, and well exposedness) and colour dissimilarity feature. Finally, the fused image is generated using the multi-resolution blending utilising pyramid decomposition. The vital advantage of the proposed technique is that it blends the multi-exposure images very well and avoids the seams effectively. The method can be used for consumer cameras as the presented technique is quite fast. Experimental results prove that the proposed method produces high dynamic range images without ghost artefacts. Furthermore, the comparison, objectively and subjectively, of the proposed technique in terms of mutual information, MEF structural similarity index, and natural image quality evaluator, with state-of-the-art techniques, shows significant improvement of the proposed scheme over existing techniques.

The presence of noise is generally believed to suppress the actual signal, and subsequently, deteriorate the quality of the signal or image. However, this study counter argues, and proposes a novel technique of enhancing variably illuminated document images through the addition of external noise, using the famous physical phenomenon of stochastic resonance. Here, the fundamental relationship between inherent parameters of a dark and low contrast image, and the intensity of external noise-induced to produce best contrast enhancement is systematically modelled using regression line parameters. With this, the amount of noise to be added is automatically predicted, and is utilised to enhance the dark and low contrast document images. The proposed enhancement technique using noise-induced (non-dynamic) stochastic resonance is computationally very efficient, unlike the existing methods which estimate the optimum value of noise to be added through a time-consuming iterative method. The efficacy of the developed technique is demonstrated through experimental results. Overall, a mathematical model for predicting the optimum noise standard deviation (NSD), and subsequently three variants of image enhancement algorithms based on global-NSD, partition-NSD and mask-NSD are the major contributions reported in this research work.

Vision-based human activity recognition (HAR) finds its application in many fields such as video surveillance, robot navigation, telecare and ambient intelligence. Most of the latest researches in the field of automated HAR based on skeleton data use depth devices such as Kinect to obtain three-dimensional (3D) skeleton information directly from the camera. Although these researches achieve high accuracy but are strictly device dependent and cannot be used for videos other than from specific cameras. Current work focuses on the use of only 2D skeletal data, extracted from videos obtained through any standard camera, for activity recognition. Appearance and motion features were extracted using 2D positions of human skeletal joints through OpenPose library. The approach was trained and tested on publically available datasets. Supervised machine learning was implemented for recognising four activity classes including sit, stand, walk and fall. Performance of five techniques including K-nearest neighbours (KNNs), support vector machine, Naive Bayes, linear discriminant and feed-forward back-propagation neural network was compared to find the best classifier for the proposed method. All techniques performed well with best results obtained through the KNN classifier.

Intervention surgery strongly requires information on the guide-wire position under the monitoring of X-ray video. Hence, the related researches such as guide-wire detecting or tracking have become widespread. However, most of the existing methods require a lot of resources for computing or large data for training since the X-ray videos have internal physicals such as anatomical skeleton contours and organs that are quite similar to a guide-wire. This work presents a practical method that only requires a moderate number of training data for detecting a guide-wire tip in an X-ray video sequence during the percutaneous coronary intervention surgery. The method applies maximally stable extremal regions (MSER) combine with modified multi-filters (region area range filter and stroke width variation filter) for region detection and local binary patterns (LBP) for guide-wire recognition. The motivation for applying MSER and LBP are the robust efficacy and the low requirement of resources. The approach evaluated 20 different sequences of X-ray videos, a total of 1295 frames. 50 selected frames were used as training templates and others to experiment. The method was successfully performed to the detecting guide-wires with p-value < 0.01 compared with conventional MSER methods, 93.7% average detection accuracy, and 21 fps average speed.

Design of low-power and area-efficient portable complementary metal–oxide–semiconductor processors for image and signal processing applications demand reduction in transistor switching and count. Adder is the fundamental block of all arithmetic operations performed in processing units. In this study, an error-tolerant parallel adder with faithful approximation is proposed that can optimise area and accuracy. In the proposed parallel adder, for n bit input and m bit adder block, least n/2m blocks are designed with approximate logic using carry by-pass addition algorithm and most n/2m blocks are designed with exact logic using carry select addition algorithm. Least significant approximate part of the adder is designed with either exact full adder (EFA) or fault-tolerant full adder (FTFA) cells. This confines the maximum error in the proposed-EFA and proposed-FTFA designs to be not more than unit bit value with weights 2^{[(n/2m)−1]m} and 2 ^{n /2}, respectively. Two different FTFA cells are proposed and implemented in the approximate blocks. The synthesis results of the proposed-EFA, proposed-FTFA1 and proposed-FTFA2 designs using Cadence Encounter with 90 nm ASIC technology for n = 16, m = 4 demonstrated an area saving of 22.3, 28.2 and 35%, respectively, when compared to the conventional counterpart.

A modified digital image processing technique is presented to accurately investigate the types of the acute lymphatic leukaemia (ALL). In this technique, three complementary steps are performed. In the first one, a colour segmentation procedure is used to obtain images including only the white blood cell. In the second step, the histogram equalisation and linear contrast stretching procedures are utilised to obtain images for the nucleus. In the third step, images for the cytoplasm only may be reconstructed from which the vacuoles may be detected. For accurate detection for ALL types, significant and discriminative parameters are introduced such as geometric shape of nucleus membrane, equivalent sizes for the nucleus and cytoplasm and their ratio when the shapes of nucleuses are regular or irregular. This method is applied to a blood smear images for real cases of ALL. To validate the present technique, a comparison is made between present results with their counterparts obtained by expert (manual) technique. Another assessment is performed by comparing the average accuracy of the present technique and the average accuracy of different image processing techniques in the literature. The assessment confirms the high efficiency of the present technique in detecting all types of ALL.

In this study, the authors propose a new loss function for denoising convolutional neural network (DnCNN) for salt-and-pepper noise (SPN). Based on the motivation of utilising the mask of SPN, firstly from the usual SPN-denoising restoration equation, the authors establish a perfect restoration condition; the restored image is precisely the clean image if this condition holds. Then they design a mask-involved loss function to encourage the network to satisfy this condition in training progress. Experimental results demonstrate that compared with general DnCNN and other state-of-the-art SPN denoising methods, DnCNN equipped with the proposed loss function involving mask (MaskDnCNN) is more effective, robust and efficient.

As one of the most popular kinds of the component substitution (CS)-based pansharpening methods, the intensity-hue-saturation (IHS) method can produce the pan-sharpened images with high spatial quality while causing some spectral distortion, mainly owing to it cannot estimate an accurate intensity image in the IHS space. To solve this issue in the IHS method, in this study, the authors propose a new pansharpening method with gradient transferring in the generalised IHS transform space, which aims at estimating a more accurate intensity image. More specifically, the novelty of the proposed method consists of building a novel variational gradient transferring model to transfer the spatial gradient information of the panchromatic image into the new intensity image as well as preserve the local spectral information from the low resolution multispectral image. Finally, they compare the proposed method with some CS methods using the Pleiades, QuickBird, and GeoEye-1 satellite datasets from both the subjective and objective aspects. Specifically, the experimental results show that the proposed method yields better pansharpening results than the other methods in terms of higher spatial and spectral qualities.

In this study, the authors propose a framework SUGAMAN (Supervised and Unified framework using Grammar and Annotation Model for Access and Navigation). SUGAMAN is a Hindi word meaning ‘easy passage from one place to another’. SUGAMAN synthesises textual description from a given floor plan image, usable by visually impaired to navigate by understanding the arrangement of rooms and furniture. It is the first framework for describing a floor plan and giving direction for obstacle-free movement within a building. The model learns five classes of room categories from 1355 room image samples under a supervised learning paradigm. These learned annotations are fed into a description synthesis framework to yield a holistic description of a floor plan image. Authors demonstrate the performance of various supervised classifiers on room learning and provided a comparative analysis of system generated and human-written descriptions. The contribution of this study includes a novel framework for description generation from document images with graphics while proposing a new feature representing the floor plans, text annotations for a publicly available data set, and an algorithm for door to door obstacle avoidance navigation. This work can be applied to areas like understanding floor plans and design of historical monuments, and retrieval.

A novel image denoising algorithm which is based on the ordering of noisy image patches into a 3D array and the application of 3D transformations on this image dependent patch cube is proposed. For a given noisy image, the authors extract all the patches with overlaps. Then, they order these patches according to a predefined similarity measure. After reordering, a possibly separable 3D transformation is applied to the reordered 3D patch cube. The transform domain coefficients are thresholded using a suitably calculated thresholding parameter. Afterwards, the proper 3D inverse transformation is applied to these coefficients. The final denoised image is generated by repositioning the processed patches to their original locations on the image canvas. The developed algorithm presents a novel and efficient combination of patch ordering and 3D transformations. The forward analysis transform as defined by this complex procedure can get restated as the application of a single tight frame. This tight frame depends on the noisy image under consideration. This novel, image dependent forward operator which employs 3D transforms results in improved denoising performance. The experimental results indicate that the proposed algorithm achieves state-of-the-art denoising results with complexity comparable to competing methods.

Welding defect detection in a radiographic image is an important topic in the field of industrial non-destructive testing. To improve the accuracy of welding defect segmentation, a local image enhancement approach is proposed. In this algorithm, the requirement of contrast enhancement is considered when extracting the weld seam and segmenting the weld defect. The whole defect detection is conducted by three procedures: image enhancement, welding seam extraction, and defect segmentation. Firstly, a method for determining the Localised Pixel Inhomogeneity Factor (LPIF) is proposed. Then, based on the results of LPIF, the Otsu method is applied to segment the welding seam and defects are, identified by region growing algorithm. The authors compared LPIF with histogram equalisation, adaptive histogram equalisation, and contrast-limited adaptive histogram equalisation algorithms and assessed its performance by using indicators such as image contrast, image definition, and edge intensity. Moreover, the authors compared the segmentation results of the enhanced defect images with the original image to further study the method's effect on weld defect segmentation. More than 70 images containing various types of defects are tested. The experimental results demonstrate that the quality of enhanced defect images is improved significantly, and has a high relative segmentation accuracy of more than 92%.