Deep Learning-Based Object Detection in Diverse Weather Conditions

Introduction

Computers are like machines which unlike humans can understand only numbers (on a high level). But the advancements in technology require computers to deal with and understand an image or video data. Thus, to create a bridge between the two, Computer Vision (CV) and Image Processing (IP) played an important part. The extremely vital field in Artificial Intelligence (AI) helps to detect objects in different weather conditions. The use of Object Detection (OD) is everywhere around like it is used for Content-based image retrieval, Surveillance Industry, Traffic tracking systems, Activity Recognition, Sentiment Analysis, Human-Computer Interaction, Document Summarization, Robot Vision, Consumer Electronics, Security Systems, Autonomous Driving, and Augmented Reality. OD is the most critical problem in CV. Also, this is the main task and is at a high level in the field of Artificial Intelligence that coexists with us into our lives. There are several OD methods (Saritha, R. R. et al., 2019) that exist. The goal of OD is to build rectangular bounding boxes around the objects and figure out what class they belong to. Scenes in images can be analyzed or understood with the help of recognition of image that comes under computer vision which is linked to OD. Elements in scenes can be understood at pixel level which is created by image segmentation whereas only a class label could be obtained via IR. The motivation behind this topic is that it is the strategy that will overdo all the physical tasks like Counting in Crowd, Cars: Self Driving, Surveillance System, and Detection of a Face in any type of weather.

OD is used in several applications and this can be customized up to various extents based on the real-time requirement. There are different ways to detect an Object like Point-based Detection and Optical Flow, out of which, the most frequent ones are Stage Detectors that can detect an Object in a Single Shot. The detection can be traditional or deep learning-based and it can be a two-stage or one stage. In one-stage detector methods, objects can be detected in a single go whereas in two-stage detectors the objects cannot be detected in a single go instead they can be detected in two phases.

Deep Learning has a great deal of potential in the Object Detection arena, and it can help us construct a complete solution. These algorithms can also be strengthened to provide predictions that are as near to the original bounding box as possible. As a result, the system will produce tighter and sharper bounding box estimations.

The Transfer Learning approach has been used in an effective way to detect an object. Like here it is not required to build and train the model from the start but to make faster progress the pre-downloaded weights can be used which are trained by someone else on the network architecture. So, this is a perfect fit that can be used as a pre-trained model to transfer the knowledge for a task like detecting an object in adverse weather conditions like snowy, hazy, rainy, foggy, etc., The computer vision research society has done a fine job of publishing a lot of data sets on the web, such as image net or coco or pascal kinds of data sets. These are all titles of different data sets that people have been posting online and that a lot of computer scholars have skilled their algorithms on. Often this training takes several weeks and may require many GPUs. A few open-source implementations of neural networks are available for download, including not only the code but also the weights, and there are many networks to choose from that have been trained on, such as the image net data set, which has thousands of different classes, so the network might have a softmax unit that outputs one of a thousand multiple categories. What else can be done is to delete the existing softmax layer and create a new softmax unit, or the first few layers can be frozen, and post-processing can be accomplished based on the structure. For example, depending on the framework, stuff like trainable parameter equals zero can be set for these earlier layers, and then a shallow softmax model can be trained from this feature vector. Let say, there is one Pretrained Model which is already trained on huge datasets, then it can be used with a small dataset for Post Processing of layers and it can be used for a customized purpose.