Custom CV Model Improves the Accuracy of Image Theft Detection Solution from 65% to 88%

The founders of PhotoStat were looking for companies specializing in Computer Vision services via Clutch and stumbled upon Maruti Techlabs' profile. A couple of calls and meetings later, it was clear that both companies aligned well, and Maruti Techlabs was signed as their Computer Vision Services partner.

These were some of the reasons Maruti Techlabs stood out to the founders:

• MarutiTech's expertise and experience in Image Recognition and Segmentation
• Our work in Object Recognition, OCR, and Process Automation
• Our expertise in refining and sorting different datasets
• High degree accuracy of the AI and ML models we have built
• Reviews from our clients on the Maruti Techlabs Clutch profile

1. Feasibility Study: 
After thoroughly understanding the client's use case and requirement, the first thing we worked on was conducting a feasibility study spanning four weeks, wherein our AI experts defined the scope of the solution and conducted a detailed analysis of the platform's current state.

During the feasibility study, our AI experts work with a sample of the image datasets (provided by the client) to determine the feasibility of the desired outcome. After creating the training dataset acquired from the client's database, we filtered, organized, and labeled the dataset to make it search-friendly. The labeled dataset then underwent meticulous quality checks like adding or removing pixels, removing noise, and sorting misclassified data.

After the data was processed, our engineers then leveled the dataset using various techniques like flipping, cropping, blurring, zooming, and compressing as required. Fundamentally, our AI experts studied, preprocessed, matched the sample data, and defined approaches they would take to build the search engine.

2. Development:

Once we completed the feasibility study and refined the training datasets, the actual development work began. With the help of training datasets, we designed a search engine, using computer vision technology, that gives similar user images for the given input image (uploaded by the artist).

Along with the similar images, the search engine also shows the percentage of similarity for each resultant image in the form of similarity distance. For easy interpretation of data, we created three buckets wherein the search engine classifies the results based on the similarity distance score. The three buckets are:  

• Exact Match
• Similar Match
• No Match

We built the image search engine as a 5-layered architecture wherein: 

1. The first layer preprocesses the images for noise removal, normalization, filtration, resizing, grayscale conversion, etc.

2. The second layer acts as the feature extractor and feature hash generator. This layer extracts the features of images in comparison. For the feature hash generator, we built a hash generation algorithm based on the features extracted for the images.

3. We built an index tree for the third layer to enable faster searching of images based on their feature vectors.

4. The incoming image bucketing is done based on similarity distance calculation. If the initial distance scoring via feature hash matching does not give a confident result (distance score more than a certain threshold value), then the fourth level of logic matches the feature vectors. And it calculates the percentage for features of user image matching in the public image to accommodate picture-in-picture similarity scoring.

5. An additional fifth layer was built for logo detection wherein a separate model was built and trained on labeled data of logo images.