Description:Field of Invention
The current invention is meant for automatic detection recognition of objects from images using deep learning and bio-inspired methods. A hybrid bio-inspired optimization method known as Particle Swarm Firefly Algorithm (PSFA) is combined with an existing deep learning model known as FRCNN. It has provision to detect multiple objects present in given image. Object detection and recognition of objects with higher level of accuracy is achieved with the current invention. It is achieved with integration of deep learning and hybrid bio-inspired method. With multiple layers’ depth using deep learning and optimization using hybrid bio-inspired method, the current invention leverages efficiency in object detection.
The objectives of this invention

This invention is designed to improve quality in object detection and recognition in computer vision applications. It is achieved by using a hybrid bio-inspired optimization method that combines firefly and particle swarm to optimize weights of FRCNN model. The current invention is evaluated and found to be more efficient than its counterparts.
Background of the invention
Computer vision applications are realized with artificial intelligence (AI). Deep learning methods such as Convolutional Neural Network (CNN) became popular for analyzing image content. As the image processing is done with higher level of accuracy with CNN, it is widely used in different real world applications. There are many existing methods that are based on CNN model for object detection. (Vipul Sharma, et.al, Journal of King Saud University , Vol. 34, No..5, pp: 1687-1699, 2022)invented FRCNN based method known as saliency guided FRCNN to detect and recognize objects. Saliency guidance is found to be an important optimization to leverage functionality of FRCNN. (Cui-Jin Li, et.al, Pattern Recognition Letters , Vol. 145, no.4, pp: 127-134, 2021) exploited FRCNN model in order to detect multiple objects from given images. Particularly, they considered images pertaining to traffic environment where object detection plays crucial role. Their methodology is based on cross-layer fusion approach which is found to be efficient in improving detection performance. (Chongqing Cao, et.al, IEEE Access, 2017) focused on detection of small objects. Small object detection is important in computer vision applications where small objects are to be detected to realize complete object detection process. (CanPeng, et.al, Pattern Recognition Letters , Vol. 140, no.4, pp: 109-115, 2020) improved FRCNN with incremental learning towards realizing object detectors. The incremental approach has potential to expedite detection process besides improving training accuracy. (Kun Wang, et.al, IEEE Access , Vol. 8, No..5, pp: 193168-193182, 2020] on the other hand investigated on object detection in images pertaining to night scenes. They combined FRCNN model and DCGAN model to improve detection accuracy. (SubramanianParvathi, et.al, Bio system Engineering , Vol. 202, No..5, pp: 119-132, 2021) used FRCNN model to detect maturity stages of coconuts presents in images that are having very complex background.
This research could have utility in agriculture applications. Studied many frameworks along with SSD, YOLO and FRCNN to recognize objects and type of vehicles. They made a comparative study among the models and found that FRCNN has potential to improve object detection accuracy. (EminGüney, et.al, SakaryaUniversty Journal of Computer and Inormation Sciences, Vol. 5, No..2, pp: 217-224, 2021) used FRCNN to solve problems in transportation domain. In other words, they used the deep learning model to detect road objects and traffic signs in traffic environment. (Fen Fang, et.al, IEEE Transactions on Image Processing, 2019) focused on elongated object detection. Their methodology combines model-driven clustering and FRCNN to realize faster and accurate detection of objects. (HaiHaung, et.al, NeuroComputing , Vol. 337, pp: 372-384,2019) exploited data augmentation along with FRCNN to detect and recognize marine organisms while FRCNN is used .(Wang Cheng,International Conference on Robots & Intelligent System, pp:204-206,2019) for detecting objects. The state of the art clearly states that FRCNN is widely used for object detection process. However, the existing methods explored little on optimizing weights of FRCNN model. The current invention proposes a hybrid bio-inspired method to achieve this.

Detailed of Prior Art
Many patent applications have been found on object detection research. A system is developed for detecting local and global objects. The global detection techniques are employed to detect global object while the local detection techniques are employed to detect local objects (US9064172B2). Object detection and also recognition on video frames is another invention found. Object detection process is made simultaneously on multiple video frames and then they rare recognized (US10699126B2). Detection of complex objects is another invention found in the patent literature. Object detection is done in one of more video frames using different techniques and methods. Particularly, complex objects that are not easier to detection are focused on this invention. Different levels of detecting objects is employed to have a systematic approach in finding all objects and localizing them using bounding boxes (US11004209B2). Object detection is made for the purpose of autonomous driving system. It has object detection process supported by a confidence score that could improve detecting performance. It has ground truth based learning approach to achieve maximum coverage in object detection (US11210537B2). There is another invention found on complex object detection in video frames. It makes use of a camera and for obtains live data from the surroundings and detects simple and complex objects (US20220268933A1).

Summary of Invention
The current invention combined deep learning model FRCNN with a hybrid bio-inspired method for leveraging object detection performance. The invented bio-inspired extension to FRCNN makes it more intelligent in making decisions. The hybrid bio-inspired method is designed to optimize weights of FRCNN model for influencing better performance. Moreover, the current invention has improved objective function besides usage of number of feature detectors to improve quality of training and detection of objects. Performance evaluation has revealed the efficiency of the current invention which has potential to solve real world problems though it is designed for object detection and recognition.
A detailed description of the invention
In this considered work a hybrid optimization technique is invented by combining particle swarm with firefly algorithm. This hybrid optimized technique is combined with faster RCNN to obtain high level of rate of accuracy.
First, Faster RCNN is trained with the help of invented weights which are predefined in neural networks. The dataset consists of images and all the images are tuned into even optimized and converted into particles. The PSO further optimizes the weights as the weights are converted as particles. After getting optimized Fireflies are generated. In which the swarm particles are replaced with Fireflies and the search position is updated with respect to intensity of light to obtain the best solution. Finally update Faster RCNN with the weights obtained after applying hybrid optimization technique. Hence the output prediction is calculated and the results are shown. The main aim is to increase the prediction accuracy rate.dimension; n number of nodes are taken as input. The training is performed in the layers and one encoded vector will be consider as output. After the process of training these weights will be
Our object detection system consists of three modules. The first generates category-independent region proposals. These proposals define the set of candidate detections available to our detector. The second module is a large convolutional neural network that extracts a fixed-length feature vector from each region. Finally, third module is we perform classification. The entire process is performed on PASCAL VOC 2007 dataset. The weights updated by using FRCNN is fed to hybrid firefly so that more accurate and better results can be obtained. The hybrid model is combination of particle swarm and firefly algorithm. The hybrid model is discussed in below section. In the faster RCNN anchors which are termed as bounding boxes are present. Here each object is surrounded with a bounding box. The trained data set consists of bounding boxes and which works well for PASCAL VOC dataset. The use of bounding boxes increases the speed of detecting the object and improves in providing high rate of accuracy. The feature extraction performed based on the objects, were all the similar features of the objects present in the bounding boxes are extracted, and regression process is performed in different layers so that major feature vectors are extracted.
Particle Swarm Firefly Algorithm (PSFA),
Both the techniques Swarms and flies are approximately near to each other. Fireflies communicate with each other, try to find the pray and identifies mates using the light pattern which gets flashed. Firefly algorithm is based on some of the ideal traits of fireflies
a. All fireflies are defined as unisex so that the fireflies get attracted irrespective of their sex.
b. Here we compared the relation between attractiveness and brightness which are proportional to each other, thus for any two blinking fireflies, as the proportionality is considered between attractiveness and brightness the distance between the flies increases gradually and the less bright one will move towards the brighter one. When there is no firefly which is brighter than the firefly takes a step towards the other fly.
c. The brightness of a firefly is affected or determined by the landscape of the objective function.
The current invention is characterized by an important contribution to the field of object detection. This contribution is a hybrid bio-inspired method that combines fire fly and particle swarm intelligence. This method is used to optimize weights of FRCNN model in order to improve object detection accuracy. When two fireflies such as x and y are considered, it is possible to have two solutions. The distance between them is computed as given in Eq. 1.
D_xy= v(?(P_x-P_y)?^2 ) (1)
Based on the value of D_xy it is possible to compute attractiveness. Thus x fly’s new position is determined as expressed in Eq. 2 and Eq. 3.
ß= ß_o e^(-?D_xy^2 ) (2)
P_xynew= P_x+ß.rand ?P_xy+ rand (3)
A random number rand is considered to have random solutions produced for x and ß_o. With regard to attractiveness, step wise computation and update is done as in Eq. 4.
?P_xy=(P_y-P_x) (4)
To achieve global best solution P_x^new= P_x if P_x isP_Gbest and P_x^new= P_xGbest^new if P_y isP_Gbest and it complies with the condition P_x^new= P_xy^newwith ?Fit?_best. With particle swarm fire fly method is improved to have optimal weights to FRCNN model. In the process different computations are involved as in Eq. 5 to Eq. 7.
D_xbest= v(?(P_x-P_Gbest)?^2 ) (5)
?P_1xy= P_Gbest-P_worst (6)
?P_2xy=(P_y-P_x+P_D1-P_D2) (7)
In the invented hybrid method fitness function is improved as in Eq. 8.
?P_xy= ?P_1xy if?FF?_x>?FF?_Npop , otherwise ?P_xy= ?P_2xy (8)
Controlling fitness function plays crucial role in the invented method. The controlling process is expressed as in Eq. 9 and Eq. 10.
?FF?_x= (?Fit?_x-?Fit?_Gbest)/?Fit?_Gbest (9)
?FF?_Npop= (?Fit?_Npop-?Fit?_Gbest)/?Fit?_Gbest (10)
In the invention a distribution function is considered and there is uniform distribution for effective swarm search leading to a solution expressed as in Eq. 11.
P_xy^new= P_x+ ßrandn?P_xy+randn (11)

In summary, this invention has two aspects involved. First, it has FRCNN to achieve object detection and the second one is the hybrid bio-inspired method that is made up of firefly and particle swarm methods. The bio-inspired method is designed to optimize performance of FRCNN method. In other words, the weights involved in the FRCNN method are subjected to optimization with the help of bio-inspired method discussed above. Thus the current invention has potential to leverage object detection accuracy.

Brief Description of Drawing
Here are a list of Figures reflecting exemplary embodiment of the current invention.
Figure 1: Overview of FRCNN for efficient object detection and recognition. FRCNN is later combined with a hybrid bio-inspired method
Figure 2: Functional flow of the current invention exploiting FR-CNN and hybrid particle swarm firefly method
Figure 3: Object detection results and corresponding Image
Figure 4: Object detection results and corresponding average precision
Detailed Description of Drawing
The current invention is meant for automatic detection recognition of objects from images using deep learning and bio-inspired methods. A hybrid bio-inspired optimization method known as Particle Swarm Firefly Algorithm (PSFA) is combined with an existing deep learning model known as FRCNN.
Figure 1 illustrates overview of FRCNN for efficient object detection and recognition. FRCNN is later combined with a hybrid bio-inspired method.On receiving input images, the deep learning model uses underlying convolution layers to acquire feature maps in order to have representation of given image. Pooling layers in the deep learning model are used to optimize feature maps. The deep learning model has underlying mechanisms to have region based approach to expedite object detection process. Feature extraction is thus made easier and faster. Feature representation with multiple techniques is another important consideration for improving accuracy in object detection. It has different modules to achieve this. First, it produces regions that are category-independent. This will help to have multiple objects to be detected in the invented system. Another module is CNN which produces feature vectors of fixed length for each region. There is another module designed for classification. The FRCNN model is integrated with a hybrid bio-inspired method for leveraging detection accuracy.
Figure 2 illustrates the functional flow of the current invention exploiting FR-CNN and hybrid particle swarm firefly method. The hybrid bio-inspired method is designed to optimize weights of FRCNN model towards higher level of detection accuracy. Swarm intelligence and firefly reflect behaviors in nature. Both are bio-inspired method that are meant for optimization. In the current invention, both are combined seamlessly to influence efficient optimization of detection process in the invented system. Fireflies involved in the invented system have systematic communication to know pray and exploit light pattern generated. Fireflies do exhibit number of traits such as having unisex to be sex-independent, there is relation between brightness and attractiveness, and improved objective function. The firefly method combined with swarm optimization has potential to improve optimization. In the current invention, the hybrid bio-inspired method could optimize weights of deep learning model.
Figure 3.illustrates object detection results and corresponding Image, it shows the accurate detection of objects such as baby. Figure 4.illustrates object detection results and corresponding average precision of experiment
5 Claims & 4 Figures , Claims:The following claims define the scope of the invention:

Claims:
1. A bio-inspired hybrid technique that combines PSO and firefly to improve the performance of the FRCNN deep learning model for object detection and recognition.
a) The FRCNN Compute weights required for object detection
b) The weights are optimized by hybrid optimization Algorithm and FRCNN performance is improved.
c) The optimization of weights has potential to leverage accuracy in object detection.
2. As per claim1, module for seamless integration that combines FRCNN and bio-inspired optimization techniques.
3. As per claim1, as part of the current invention, the Particle Swarm Firefly Algorithm (PSFA) method makes use of two bio-inspired techniques for enhanced optimization
4. As per claim1, a better fitness function is defined to increase the method's intelligence
5. As per claim1, a feature selection method along with different feature representations for efficient training of deep learning mode involved in the current invention.