Automatic detection of plant diseases has emerged as a challenging field in the last decade. Computer vision-based advancements have helped in timely and accurate identification of diseases, making possible an appropriate treatment and hence ensuring an increased yield. Diseases attack in different formations on a plant; the most severe being multiple diseases appearing on a single leaf. Moreover, as various diseases progress, they generate similar-looking symptoms making the task of identification further difficult. This work addresses these two problems with the help of an improved YOLOX model. We propose a modified Spatial Pyramid Pooling (SPP) layer to effectively extract relevant features at various scales from the training data. It is achieved by concatenating multilevel features pooled from smaller to larger scales. To enhance generalization capability of the design, various skip connections are also introduced. To improve the network convergence and detection accuracy, α _IoU based regression loss function was employed. A dataset composed of 1, 112 cotton plant images with co-occurring diseases along with their progressive severity levels was collected from the Southern Punjab region of Pakistan. Apart from healthy images, the dataset comprises of 3 severity stages of cotton leaf curl with co-occurring cotton sooty mold stress on single leaf image. Experimental results revealed that our proposed improved SPP-based YOLOX-s model achieved 73.13% mAP on our self-collected dataset and achieved 3.27% better test accuracy than original YOLOX model.

Using deep learning-based tools in the field of agriculture for the automatic detection of plant leaf diseases has been in place for many years. However, optimizing their use in the specific background of the agriculture field, in the presence of other leaves and the soil, is still an open challenge. This work presents a deep learning model based on YOLOv6s that incorporates (1) Gaussian error linear unit in the backbone, (2) efficient channel attention in the basic RepBlock, and (3) SCYLLA-Intersection Over Union (SIOU) loss function to improve the detection accuracy of the base model in real-field background conditions. Experiments were carried out on a self-collected dataset containing 3305 real-field images of cotton, wheat, and mango (healthy and diseased) leaves. The results show that the proposed model outperformed many state-of-the-art and recent models, including the base YOLOv6s, in terms of detection accuracy. It was also found that this improvement was achieved without any significant increase in the computational cost. Hence, the proposed model stood out as an effective technique to detect plant leaf diseases in real-field conditions without any increased computational burden.

Traffic sign detection and recognition has been a topic of research for at least the last two decades. Efforts are being made to reliably detect candidate traffic signs in natural uncontrolled environment and to recognize their contents. For detection, a large proportion of relevant literature discusses color based segmentation by either sticking to a predefined color space (e.g., RGB, HSI, YCbCr etc.) or make use of empirically selected subset of eigen space to achieve partially data dependent segmentation. Since, the input RGB data for various color classes and the background is not linearly separable, none of the existing methods guarantee to achieve complete separation among pixels corresponding to traffic signs and the background objects. To tackle this problem, we propose a completely data driven segmentation technique that adaptively selects an optimized color space based on available training data. To recognize the contents of potential traffic signs, we present a hybrid spatio-frequency radial feature extraction technique with an emphasis on the regions containing useful information. We explore the energy compaction property of steerable discrete cosine transform for feature extraction and augment it with well known circular histogram of oriented gradients in a pyramid. Using our proposed method, experiments on (1) German Traffic Sign Detection Benchmark, (2) our self collected dataset and on a (3) hand crafted version of the combination of the two provide competitive performance compared to various latest and state of the art methods by achieving up to 0.978 precision and 0.98 recall values at an expense of only an insignificant additional computational cost. The method also obtained 0.81 precision on traffic signs partially occluded by other objects.

The utilization of deep learning-based models for automatic plant leaf disease detection has been established for many years. Such methods have been successfully integrated in the agriculture domain, aiding the swift and accurate identification of various diseases. However, the unavailability of annotated data, the variability of systems, and the lack of an efficient model for real-time use remain unresolved challenges. The goal of this work was to develop a deep learning-based model for crop disease detection and recognition system for real-field scenarios. For this, we trained lightweight versions of the YOLOv5, YOLOv7, YOLOv8 and compared their detection performance. Experiments were carried out on a self-collected dataset containing 3136 real-field images of apples ( healthy and diseased ) and 567 images of PlantDoc dataset. Results revealed that the prediction accuracy of YOLOv8 was superior to others on AdamW optimizer. The results were further validated by deploying it on Raspberry Pi 4.

Computer vision based traffic sign detection and recognition is an active field of research but the task becomes challenging when the sign of interest is partially occluded by nearby objects like a tree, pole or vehicle. Another difficulty posed especially in the developing countries is the lost colors problem that arises due to aging and poor maintenance. This work presents an automatic technique that focuses on visible parts only and suppresses occluded portions. Features are collected using a convolutional neural network inspired invariant feature extraction technique augmented with feature interaction based dimensionality reduction. Further, with the use of dynamic parameter estimation, an adaptive system for continuous learning is also proposed. Since the effect of partial occlusion has not been thoroughly studied, there is no benchmark database available for this purpose. We have prepared two datasets by combining originally and synthetically occluded images taken from field surveys and from famous GTSRB database. Experiments revealed that our technique outperforms state of the art recognition methods previously used for visible and occluded signs by obtaining 0.81 precision and 0.79 recall values on the average. The proposed method also shows a remarkably low error rate as the amount of occlusion is increased.