Multi-class object detection system using hybrid convolutional neural network architecture

Object detection in computer vision has been a significant research area for the past decade. Identifying objects with multiple classes from an image has attracted great attention because it can effectively classify and detect the image. A multi-class object detection system from a video or image is quite challenging because of the errors obtained by the location classification process. Our proposed system generalized a hybrid convolutional neural network (H-CNN) model is used to realize the user object from an image. The proposed work integrates pre-processing, object localization, feature extraction and classification. First, the input image is pre-processed with Gaussian filtering to remove noise and improve the image quality. After completing the pre-processing procedure, it is subjected to object localization. Here the object in the image is localized using Grid Guided Localization (GGL). In the feature extraction phase, the model would be pre-trained with AlexNet. Here the AlexNet are generalized as fully connected (FC) layers. Finally, the Softmax layer in the AlexNet architecture is replaced by SVR (Support Vector Regression), which acts as a classifier for identifying the object class. The classification loss is minimized using the Improved Grey Wolf (IGW) optimization algorithm. Thus, the H-CNN model can quickly classify and label the objects from images. It also offers improved classification performance in managing effective training time. The proposed work will be implemented in PYTHON. Therefore, the model would be built using various datasets such as MIT-67, PASCAL VOC2010, MS (Microsoft)-COCO, and MSRC to effectively train and classify the object. The proposed H-CNN achieved improved results with MIT-67 (96.02%), PASCAL VOC2010 (95.04%), MSRC (97.37%), and MS COCO (94.53%). The results obtained by H-CNN proved that the excluded result of Mean Average Precision (mAP), Precision, Accuracy, Recall values ​​and F1-Score achieved better results than with recently developed works such as YOLO-fine, EfficientDet, YOLOv4, RetinaNet, GCNet and HRNet architectures.