نشریه ماشین بینایی و پردازش تصویر
سال دهم شماره 4 (زمستان 1402)

Fruit detection due to different lighting conditions, occlusion and overlap is a challenging task in machine vision-based harvesting robots. The aim of this article is to improve the accuracy-speed trade-off in apple fruit detection in the vision system of agricultural harvesting robots. Considering the recent applications of attention modules in the field of object detection, we have proposed a new architecture of YOLOv5 network in which the ECA channel attention module is replaced by the C3 module in the backbone of the network. Despite reducing the number of network parameters, the ECA module has not had a significant effect on the detection efficiency, and by increasing the speed by 22% compared to the YOLOv5 Nano version, it has been able to establish a better trade-off between accuracy and speed. To evaluate the proposed architecture, three datasets KFuji, MinneApple and ACFR are used in the training and testing phase, and in the case that the training and testing databases are not the same, the transfer learning method is used to improve the test results. In the case where the training and test data are the same, the use of the proposed architecture leads to a relative improvement of the trade-off by 21.2% compared to the C3 module. In the case of transfer learning where the training and test data are not the same, a relative improvement of 18% in the trade-off has been achieved.

Generative models of shapes for 2D boundaries have applications in object detection and inference from 2D images.We investigate how to learn this generative model from a training set of shape functions.The quality of the correspondence establishment significantly affects the quality of the shape models.A state-of-the-art approach for establishing correspondence is to define a regularized empirical risk for generative models, and by minimizing this risk, the correspondence between shapes is determined.The choice of the regularization parameters of the risk has a significant effect on the quality of the correspondence.In this article, by estimating the effective dimension of the principal component analysis model and using the entropy estimation of eigenvalues algorithm, we consider the effect of error variance in determining the regularization parameter for correspondence establishment.Using Our proposed algorithm leads to the following improvements in the correspondence establishment for shape models of the objects that exist in JSRT chest radiography images: 0.5 mm specificity improvement, and training time reduction from 600 seconds to 300 seconds, compared to the minimum description length method.Moreover, the specificity of the correspondence established by our proposed method is better than that established by experts' manual landmarks in terms of specificity.

This research highlights the potential of computer vision and machine learning algorithms to enhance weed control systems and decrease herbicide use in agriculture. The development of an efficient and cost-effective smart sprayer system has the capacity to not only benefit farmers financially, but also mitigate the environmental impact of herbicide application. Further investigation could explore the feasibility and practicality of implementing such systems on a larger scale in diverse crop fields. The recognition of weeds and crops based on their appearance characteristics is crucial for effective weed control systems, and the proposed BOVW algorithm demonstrated a high level of accuracy, reliability, and sensitivity in distinguishing between common weed species and sugar beet crops. The smart sprayer system, incorporating the BOVW algorithm, exhibited a high level of precision in identifying the product from weed species online. Notably, the provided sprayer system significantly reduced herbicide consumption by 78.93% and 69.38% in the best and worst mode of spraying, respectively. The findings suggest that the variable rate mode utilizing the BOVW detection algorithm represents the optimal mode of operation for the smart sprayer system.

Medical image segmentation is one of the most important steps in medical image analysis to improve diagnosis and findings. One of the most common segmentation methods in deep learning is the use of Unet networks. The presence of overlapping layers in the Unet decoder does not allow extracting information from deeper layers. Also, due to the limited range of the received field of convolution cores, long-range information and dependencies are not considered well. In this article, our goal is to place a structure in the area between encoder and decoder in the Unet model in order to fill the semantic gap between the encoder and decoder area and better extract features by paying attention to local and global features. This model makes the target region more prominent in different medical datasets. We have conducted our experiment on 6 medical data sets, and the results obtained in two evaluation criteria, Dice and Iou, show that our proposed model has better results than Unet and based methods.

Photometry is a well-known method for 3D reconstruction of objects using images taken in different lighting conditions. In this method, by knowing the light sources' direction, the normal vectors of the surface are recovered in a dense grid through the intensities recorded in the captured images. Each normal vector is then converted to the height difference in two orthogonal directions, and the simultaneous estimation of the heights for the dense grid is done by solving a system of linear, overdetermined and inconsistent equations. The miss-alignment of the coordinate system represents normal vectors and the dense grid frame of 3D reconstruction causes a systematic error in the estimation of the gridded heights map. Photometric self-calibration methods for determining the light sources’ direction are one of the causes of miss-alignments in object and surface normal vectors coordinate systems. In this paper, a sequential and iterative process is proposed to estimate and perform an appropriate rotation to the surface normal vectors. In each iteration of this method, a portion of the necessary rotation is identified in order to parallelize of the two object coordinate systems and surface normal vectors through fitting a geometric transformation to the estimated residuals of the 3D reconstruction process. The results of using the proposed method in various experiments have demonstrated a noticeable improvement in the precision and accuracy of 3D reconstruction.

One of the ways to diagnose eye diseases is to examine the Retinal Fundus (RF) images by a specialist. But in patients with cataract disease, it is very difficult to diagnose other diseases due to the blurring of the RF images. The purpose of this article is to present a method based on deep learning to increase the accuracy of diagnosing common eye diseases in the presence of cataracts and other retinal lesions. In the proposed method, in order to solve the problem of mutual overlap between eye diseases, which leads to incorrect diagnosis of the disease, the weighting technique is used in the training of the proposed model in order to increase the detection capability of the deep learning network. Also, due to the limitation in the number of images containing cataracts, including other eye diseases for training the deep learning network, various models of destruction of RF images are used to artificially simulate the images of cataracts and some retinal lesions. The results of the evaluation on the reference databases of the RF images suffering of cataract show that the proposed algorithm has been able to achieve diagnosis accuracy value of 80, 82, 79, 81, 80 and 65, respectively, for Age-Related Macular Degeneration (ARMD), Myopia (MYA), Tessellation (TSLN), Glaucoma (GL), Neovascularization (NE), and Diabetic Retinopathy (DR).